diff --git "a/data/multi_image.jsonl" "b/data/multi_image.jsonl"
new file mode 100644--- /dev/null
+++ "b/data/multi_image.jsonl"
@@ -0,0 +1,329 @@
+{"index": 0, "id": "bc7a5880-cbe3-443e-8e6e-48d5b301201f", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The slide, styled with a clean white background and titled **Visualization Experiments** in large black text at the top, presents two primary figures. On the left, a grid of bird photographs and heatmaps is displayed, labeled below with **(a) Input**, **(b) Lcls**, **(c) + Lreg**, **(d) Lcdm + Lreg**, and **$$ (e) + L_{crm} $$**. Beneath this visual grid, the text reads: **Figure 4. Visualization of vision prompt based on classification-based evaluator (b)(c) and our vision-language evaluator (d)(e), respectively. +L means that we successively add this constraint, i.e., $ +L_{reg} = L_{cls} + L_{reg} + L_{crm} = L_{cdm} + L_{reg} + L_{crm} $**. To the right of this figure are two heatmaps positioned side-by-side. The left heatmap, which corresponds to sub-image 2, is labeled **(a) CUB-200-2011 Train Set**, while the right heatmap, corresponding to sub-image 3, is labeled **(b) CUB-200-2011 Test Set**. Below these charts, the caption states: **Figure 5. The nearest description for text prompt learned by PLEor, with their similarity shown in grids.**\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/bc7a5880-cbe3-443e-8e6e-48d5b301201f_input_0.jpg", "images/multi_image/bc7a5880-cbe3-443e-8e6e-48d5b301201f_input_1.jpg", "images/multi_image/bc7a5880-cbe3-443e-8e6e-48d5b301201f_input_2.jpg"], "target_image": "images/multi_image/bc7a5880-cbe3-443e-8e6e-48d5b301201f_target.png", "width": 1920, "height": 1080}
+{"index": 1, "id": "13276d60-513b-4ad8-9b2c-0d47fcfc7438", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide features a blue banner at the top with the text \"Masked Image Training & Joint Image-Label Generation\". Below this, the slide is divided into two sections. The left side is titled \"Masked-image training (ADM codebase on ImageNet)\" and includes a table (sub-image 1) listing various metrics such as IS↑, FID↓, sFID↓, Precision↑, and Recall↑ for models like \"ADM*\", \"ADM*_g\", \"ADM\", \"ADM_g\", \"MT-Diffusion\", \"MT-Diffusion_g\", \"MT-Diffusion^f\", and \"MT-Diffusion^f_g\" along with their corresponding numerical values. On the right, the section is titled \"Joint image-label generation ImageNet\" and contains a line chart (sub-image 2) plotting \"Validation accuracy\" against \"Finetuned iteration\" from 0 to 6000. The chart displays four curves corresponding to the legend items: \"MT-Diffusion: Top-1 Acc.\", \"Supervised Pretrained: Top-1 Acc.\", \"MT-Diffusion: Top-5 Acc.\", and \"Supervised Pretrained: Top-5 Acc.\"\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/13276d60-513b-4ad8-9b2c-0d47fcfc7438_input_0.jpg", "images/multi_image/13276d60-513b-4ad8-9b2c-0d47fcfc7438_input_1.jpg"], "target_image": "images/multi_image/13276d60-513b-4ad8-9b2c-0d47fcfc7438_target.png", "width": 2880, "height": 1620}
+{"index": 2, "id": "ad873a61-83cc-4739-804a-fbbe16cf8782", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, set against a clean white background, introduces a conceptual framework titled in large, dark red serif font at the very top: **Idea: attention with optimal transport and kernel methods**. Below this header, the slide is organized into two numbered points in black text that outline the core methodology. The first point reads: **1. Non-linear layer: we use a parametrized kernel embedding in the fashion of [Chen et al., 2019a] (similar to a 1-D CNN).** The second point follows immediately: **2. Pooling: similar elements are pooled together. The measure of similarity is the optimal transport plan between the input set x and a learned reference z $\\in \\mathbb{R}^{p \\times d}$**, with specific emphasis placed on the bolded words **pooled**, **optimal transport plan**, and **learned reference**.\n\nDominating the central and lower portion of the slide are two sub-images arranged side-by-side. The first sub-image (sub-image 1) occupies the lower-left quadrant and visually depicts the transport process. It displays several blue circles labeled with input variables like $x_n$, $x_1$, and $x_2$ positioned above green circles representing reference points $z_1$ and $z_p$. Thin black lines connect the blue input nodes to the green reference nodes, annotated with labels such as $P_{11}$, $P_{n1}$, and $P_{2p}$ to represent the transport plan. To its right, the second sub-image (sub-image 2) illustrates the feature aggregation in the lower-right quadrant. It shows a row of blue rectangular blocks labeled with kernel embeddings like $\\varphi(x_1)$ and $\\varphi(x_2)$ above a corresponding row of green rectangular blocks labeled $\\Phi_z(x)_1$ through $\\Phi_z(x)_p$. Arrows crossing between the blue and green blocks, again labeled with $P_{11}$, $P_{n1}$, and $P_{2p}$, demonstrate the aggregation flow.\n\nDirectly beneath these graphical representations sits a caption in dark red text that explains the visualization: **Figure 2: The input point cloud x is transported onto the reference z = (z$_{1}$, ..., z$_{p}$) (left), yielding the optimal transport plan P$_{\\kappa}$(x, z) used to aggregate the embedded features and form $\\Phi_{z}$(x) (right).** At the very bottom of the slide, separated by a thin line, is a footer containing author credits and the paper title in a smaller, gray sans-serif font: **G. Mialon*, D. Chen*, A. d’Aspremont and J. Mairal** on the left, followed by **A Trainable OT Embedding for Feature Aggregation**, and the slide number **4 / 9** aligned to the far right.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/ad873a61-83cc-4739-804a-fbbe16cf8782_input_0.jpg", "images/multi_image/ad873a61-83cc-4739-804a-fbbe16cf8782_input_1.jpg"], "target_image": "images/multi_image/ad873a61-83cc-4739-804a-fbbe16cf8782_target.png", "width": 2552, "height": 1430}
+{"index": 3, "id": "f3fde8d2-6323-494b-ae88-819c8a449bdd", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide is titled \"Practical Implications\" at the top, followed by the sentence \"Removal-based feature attributions pass sanity checks that perturb model weights\". Beneath this text is a grid of six line charts arranged in two rows and three columns, positioned where sub-image 1 is located. The top row of charts features a \"Pearson Correlation\" y-axis, while the bottom row features a \"Rank Correlation\" y-axis. Each chart displays lines for methods including \"Occlusion,\" \"Shapley,\" \"LIME,\" and others, with the x-axis labeled with \"Original,\" \"Linear 3,\" \"Linear 2,\" \"Linear 1,\" and \"Linear 0.\" A legend at the bottom of the slide identifies the specific lines used in the visualizations.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/f3fde8d2-6323-494b-ae88-819c8a449bdd_input_0.jpg"], "target_image": "images/multi_image/f3fde8d2-6323-494b-ae88-819c8a449bdd_target.png", "width": 2880, "height": 1620}
+{"index": 4, "id": "63757db0-2f60-4ec1-84bd-d71a5ed663c3", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide is titled \"Method\" in the upper left corner. In the top right, a QR code is displayed above the text \"Project page.\" The main visual content is a technical flowchart diagram shown in sub-image 2, which occupies the central area of the slide. This diagram details a process involving a \"Source $I_s$\" and a \"Target $I_t$\" on the left, feeding into a \"Displacement estimator $f_\\theta$\" and a \"Scoring estimator $g_\\phi$.\" The flow continues to a \"Displacement $\\hat{u}$\" and a \"Scoring map $\\hat{S}$,\" eventually leading to a \"Spatial Transform\" which outputs a \"Warped source $I_s(x + \\hat{u})$.\" The bottom section of the flowchart illustrates an \"Adaptive weighting\" step resulting in a \"Re-weighted error,\" with dashed lines indicating update rules: \"Update $\\theta = \\theta - \\eta \\frac{\\partial \\mathcal{L}_{dc}}{\\partial \\theta}$\" and \"Update $\\phi = \\phi - \\eta \\frac{\\partial \\mathcal{L}_{sc}}{\\partial \\phi}$.\" Sub-image 1 is a duplicate of the QR code found in the top-right corner of the slide.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/63757db0-2f60-4ec1-84bd-d71a5ed663c3_input_0.jpg", "images/multi_image/63757db0-2f60-4ec1-84bd-d71a5ed663c3_input_1.jpg"], "target_image": "images/multi_image/63757db0-2f60-4ec1-84bd-d71a5ed663c3_target.png", "width": 1440, "height": 810}
+{"index": 5, "id": "abc9bd6e-4b85-4204-93b3-1eb0e518a2b2", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide, set against a white background with the Google Research logo in the bottom right, is titled \"More Results (Reshuffle during a roll-out)\" at the top. Below the title, a descriptive paragraph explains the methodology: \"Table 5: Reshuffle observations during a roll-out. In each test episode, we reshuffle the observations every t steps. For CartPole, we test for 1000 episodes because of its larger task variance. For the other tasks, we report mean and standard deviation from 100 tests. All environments except for Atari Pong have a hard limit of 1000 time steps per episode. In Atari Pong, while the maximum length of an episode does not exist, we observed that an episode usually lasts for around 2500 time steps.\" In the center, a blue arrow points upward toward sub-image 1, which is positioned in the middle of the slide and displays a table with columns for \"CartPole,\" \"PyBullet Ant,\" \"Atari Pong,\" and \"CarRacing.\" Beneath this table, a conclusion is written in blue text: \"Scores drop as we increase the re-shuffle frequency due to the extra time spent on input identification.\"\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/abc9bd6e-4b85-4204-93b3-1eb0e518a2b2_input_0.jpg"], "target_image": "images/multi_image/abc9bd6e-4b85-4204-93b3-1eb0e518a2b2_target.png", "width": 2880, "height": 1620}
+{"index": 6, "id": "41e7d3a6-f33f-4f54-8631-197ea6a8b409", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "Titled \"1-min Highlight: Motivation Learns a Domain-invariant Live/Spoof Classifier,\" this presentation slide is split into two comparative diagrams. The layout features a top header followed by two side-by-side sections illustrating different approaches to classification.\n\nOn the left, \"sub-image 1\" is positioned in the upper-left quadrant, depicting a transition between Live and Spoof categories across two different domains. Directly below this visual, bullet points state that the \"Domain signal is ignored,\" with the word \"ignored\" highlighted in red, and that the \"Live-to-spoof transition is inconsistent,\" with \"inconsistent\" also in red. Underneath these points, the text identifies these as \"Common Solutions: SSDG (CVPR 2020) / SSAN (CVPR 2022).\"\n\nOn the right side, \"sub-image 2\" is located in the upper-right area, showing a similar Live versus Spoof classification but with aligned arrows indicating a different relationship between the domains. The accompanying text below notes that the \"Domain signal is leveraged,\" highlighted in green, and the \"Live-to-spoof transition is aligned,\" also in green. This approach is labeled as \"SAFAS (Our Solution).\"\n\nAt the very bottom of the slide, two citation lines are provided: \"- Jia et al., 'Single-side domain generalization for face anti-spoofing,' CVPR 2020.\" and \"- Wang et al., 'Domain generalization via shuffled style assembly for face anti-spoofing,' CVPR 2022.\"\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/41e7d3a6-f33f-4f54-8631-197ea6a8b409_input_0.jpg", "images/multi_image/41e7d3a6-f33f-4f54-8631-197ea6a8b409_input_1.jpg"], "target_image": "images/multi_image/41e7d3a6-f33f-4f54-8631-197ea6a8b409_target.png", "width": 3840, "height": 2160}
+{"index": 7, "id": "7e5bb70a-bd66-4474-ad11-c1823d0a7376", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide, numbered \"4\" in the top-left corner within a blue banner, is titled \"Category query attention.\" Below the title, the text states, \"Each query learns the semantics of its category.\" To the right of this statement are two images: the upper one is labeled \"hold\" and depicts a baseball player, while the one below it is labeled \"break\" showing a similar scene but with visual markers indicating a fractured bat. Further down on the left side, another bullet point reads, \"Each query is adaptive to different images.\" Below this text, a row of two images illustrates this concept: the first shows a baseball player with the label \"hold,\" and the second, positioned to its right, displays police officers on horseback also labeled \"hold.\"\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/7e5bb70a-bd66-4474-ad11-c1823d0a7376_input_0.jpg", "images/multi_image/7e5bb70a-bd66-4474-ad11-c1823d0a7376_input_1.jpg", "images/multi_image/7e5bb70a-bd66-4474-ad11-c1823d0a7376_input_2.jpg", "images/multi_image/7e5bb70a-bd66-4474-ad11-c1823d0a7376_input_3.jpg"], "target_image": "images/multi_image/7e5bb70a-bd66-4474-ad11-c1823d0a7376_target.png", "width": 1920, "height": 1080}
+{"index": 8, "id": "e0fdaf0d-2ad9-46b0-a158-8e673864bf4a", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, numbered \"4\" in the bottom right corner, features a clean white background and focuses on a specific problem in machine learning. At the top left, the bold black title reads \"The Main Challenge of Offline RL\", underlined by a solid thin blue horizontal line that extends across the width of the slide. Directly beneath the title is a single bullet point containing the phrase \"Usual learning curve of offline RL\".\n\nThe central area of the slide is dominated by a line chart illustrating the discrepancy between estimated and actual performance over time. The vertical y-axis is labeled \"Reward\" and ranges from 20 to 120, with tick marks at intervals of 20 (20, 40, 60, 80, 100, 120). The horizontal x-axis is labeled \"total number of policy updates\" at the bottom center. The chart displays two distinct data series identified by a legend in the upper right corner: a green line representing \"real reward\" and an orange line representing \"estimated reward\".\n\nThe orange \"estimated reward\" line begins near a reward value of 30, rises steadily as the policy updates increase, and then flattens out into a plateau around the value of 90, suggesting a consistent overestimation of success. In contrast, the green \"real reward\" line also starts near 25 and rises, peaking at approximately 65 before significantly declining back down towards the bottom of the chart. This divergence highlights the core issue. Above the chart, a horizontal black arrow points from left to right, accompanied by the text \"$\\pi$ deviates from $\\mu$\", indicating that the learned policy ($\\pi$) is moving away from the behavior policy ($\\mu$) represented in the data distribution. The overall visual narrative emphasizes that while the algorithm believes it is improving (orange line), its actual performance deteriorates (green line) due to distributional shift.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/e0fdaf0d-2ad9-46b0-a158-8e673864bf4a_input_0.jpg"], "target_image": "images/multi_image/e0fdaf0d-2ad9-46b0-a158-8e673864bf4a_target.png", "width": 2540, "height": 1434}
+{"index": 9, "id": "4bfeffaf-f9d4-4688-9111-b170e9d68e43", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide is titled Language Generation as a Token-Level POMDP at the top. Directly underneath, two bullet points explain the concept: The agent’s observation is a history of tokens. and The action space is the set of possible next tokens in the vocabulary. sub-image 1 is positioned in the lower half of the slide, showing a central box labeled Transformer Language Model. Below this box is a dashed area labeled S6 containing a sequence of tokens including <a>, Hi, !, </a>, How, are, you, ?, </s>, I’m, and well. An arrow points upward from the sequence to the model, and another arrow points from the model to a green box labeled </a> next to the notation a6.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/4bfeffaf-f9d4-4688-9111-b170e9d68e43_input_0.jpg"], "target_image": "images/multi_image/4bfeffaf-f9d4-4688-9111-b170e9d68e43_target.png", "width": 1440, "height": 810}
+{"index": 10, "id": "990d0082-9870-4066-a44c-4e28de378c85", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide, titled \"Background,\" opens with the statement: \"Training large-scale generative models has emerged as the dominant approach in NLP, vision, etc.\" displayed prominently at the top. The slide is divided into two main visual sections below this headline. On the left, **sub-image 1** depicts a multi-layered neural network architecture processing text tokens, with citations referencing \"Thoppilan, et al. 'LaMDa' (2022)\" placed at the bottom. On the right, **sub-image 2** illustrates a diagram connecting a \"Transformer Encoder\" and \"Transformer Decoder\" to an image generation process involving \"ViT-VQGAN,\" accompanied by an input caption \"Two dogs running in a field\" and citations for \"Yu, et al. 'Parti' (2022).\" The Google logo is visible in the bottom right corner.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/990d0082-9870-4066-a44c-4e28de378c85_input_0.jpg", "images/multi_image/990d0082-9870-4066-a44c-4e28de378c85_input_1.jpg"], "target_image": "images/multi_image/990d0082-9870-4066-a44c-4e28de378c85_target.png", "width": 2880, "height": 1620}
+{"index": 11, "id": "a7810be6-7cc6-46be-99b5-bbecfc883930", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide is titled **Sparse Winning Tickets show ‘superior’ performance in low-data regimes**. It features a grid of six line charts, labeled as **sub-image 1** and positioned centrally, which visualize **Classification Accuracy (in %)** on the y-axis against **Network Density (in %)** on the x-axis across various training data percentages. On the upper left, a text box points to the 100% train data chart stating, **At higher data sizes, WT exist at ~(75-50%) density**. Conversely, a box on the lower left points to the 10% train data chart reading, **At lower data sizes, WT exist at ~(10-3%) density**. Finally, a text box at the bottom right connects to the 1% train data chart to note, **At least data sizes, strictly increasing trend between network performance and density**.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/a7810be6-7cc6-46be-99b5-bbecfc883930_input_0.jpg"], "target_image": "images/multi_image/a7810be6-7cc6-46be-99b5-bbecfc883930_target.png", "width": 1440, "height": 810}
+{"index": 12, "id": "f6249c90-fd42-45b4-8792-6129a10060d4", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide is titled \"Adversarial Examples\" at the top. Below the title, the text introduces the \"General idea: An adversarial example is constructed from a natural example drawn from a distribution D by adding a perturbation.\" In the center of the slide, a diagram (sub-image 1) shows a large light-blue circle intersected by a diagonal line labeled with the letter h. Inside the circle, a small blue dot is connected by a pink arrow to a purple square, situated just to the right of the dividing line. Below this visual, a legend defines the variables used in the diagram: \"c: target concept\" is shown in red text, followed by \"h: hypothesis\" in blue, and \"ρ: robustness parameter (adversary's perturbation budget)\" in purple.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/f6249c90-fd42-45b4-8792-6129a10060d4_input_0.jpg"], "target_image": "images/multi_image/f6249c90-fd42-45b4-8792-6129a10060d4_target.png", "width": 726, "height": 545}
+{"index": 13, "id": "b1526e2b-57b4-4624-bebe-33e398722cb4", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide is titled \"Related Works,\" displayed in white text within a blue banner at the top. The content below is organized into two main bullet points. The first, titled \"Formulation using regularization,\" lists three categories: \"structured sparsity-inducing norms (Friedman et al. (2010); Huang et al. (2011); Zhao et al. (2009); Simon et al. (2013); Tibshirani (1996); Bach et al. (2012); Kim & Xing (2012); Liu & Ye (2010); Rapaport et al. (2008); Zheng et al. (2018); Yuan et al. (2011); Jenatton et al. (2011)),\" \"Submodular set-functions (Bach (2010)),\" and \"Convex relaxation of linear matrix inequalities and combinatorial penalties (El Halabi & Cevher (2015); Halabi et al. (2018)).\" The second main section is labeled \"Formulation using constraints\" and introduces sub-image 1, positioned in the lower half of the slide. This sub-image features a comparative table with a black header row. The left column is labeled \"Pilanci et al. (2015) --- Our special case\" and the right column is labeled \"Our work,\" each containing mathematical formulations where the term $\\|w\\|_0 \\leq k$ is highlighted with a red box. The slide number 9 appears in the bottom right corner.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/b1526e2b-57b4-4624-bebe-33e398722cb4_input_0.jpg"], "target_image": "images/multi_image/b1526e2b-57b4-4624-bebe-33e398722cb4_target.png", "width": 1440, "height": 810}
+{"index": 14, "id": "b3fbe2c4-c718-476b-bc44-ee9dccd4b6b2", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide, titled \"(Visual) Sim2Real Transfer\" with the subtitle \"Reliability: Calibrated Sim2Real Predictions,\" features a top navigation bar indicating steps for \"Improving Generalization,\" \"Improving Reliability: Sim2Real,\" and \"Measuring Robustness.\" Below the title, the layout splits into two main sections: \"Performance\" on the left and \"(Mispredictions) Calibration\" on the right. Under the \"Performance\" heading, a bar chart labeled \"DAFormer\" is shown (sub-image 1 is positioned here), displaying values of 53.0 for \"Src-Only\" in \"Sim,\" 65.7 for \"UDA (EntMin)\" in \"Sim2Real,\" and 75.6 for \"UDA (HRDA)\" in \"Sim2Real,\" with the y-axis labeled \"GTAV → City. (mIoU).\" Below this chart, the text notes \"(Desirable) Better Sim2Real adaptation methods have improved performance.\" To the right, under the \"(Mispredictions) Calibration\" heading, another \"DAFormer\" bar chart appears (sub-image 2 is positioned here), showing values of 73.5 for \"Src-Only,\" 83.0 for \"UDA (EntMin),\" and 89.7 for \"UDA (HRDA),\" with the y-axis labeled \"Overconfidence.\" Beneath this second chart, the text reads \"(Undesirable) Better Sim2Real adaptation methods make overconfident mistakes.\" Finally, to the far right of the calibration chart, a red text box states \"Models routinely overestimate their capabilities.\"\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/b3fbe2c4-c718-476b-bc44-ee9dccd4b6b2_input_0.jpg", "images/multi_image/b3fbe2c4-c718-476b-bc44-ee9dccd4b6b2_input_1.jpg"], "target_image": "images/multi_image/b3fbe2c4-c718-476b-bc44-ee9dccd4b6b2_target.png", "width": 3840, "height": 2160}
+{"index": 15, "id": "7b18a22a-65c2-44ec-9dc3-f9abb5a5aa56", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide opens with the central question, \"Why do we need unsupervised domain adaptation for object detectors?\", displayed prominently at the top with the word \"unsupervised\" highlighted in green. Immediately below, the slide introduces a specific problem under the heading, \"Challenge 1: Object-level labels are expensive or even unavailable.\" The layout features a large photo of a golden retriever running on grass to the left, which is labeled as \"sub-image 1.\" To the right of this image, the text explains that this represents \"Image-level annotation,\" specifically identified by the label \"Dog\" quoted beneath it. The slide uses a clean white background with standard black text to contrast the green highlights.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/7b18a22a-65c2-44ec-9dc3-f9abb5a5aa56_input_0.jpg"], "target_image": "images/multi_image/7b18a22a-65c2-44ec-9dc3-f9abb5a5aa56_target.png", "width": 1920, "height": 1080}
+{"index": 16, "id": "bf6044d8-6d98-4015-ad31-59fedb1c633d", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide, titled **A Closer Look at Re-sampling** in large black font at the top, includes logos for LAMDA, University of Alberta, and Huawei in the upper-right corner. Below the title, the first bullet point reads, **Re-sampling is sensitive to irrelevant contexts**. Centered in the slide is **sub-image 1**, which displays a comparative grid of **Head Classes** and **Tail Classes** (specifically bear, fox, wolf, bicycle, bus, and train) alongside Grad-CAM visualizations for **Input**, **Uniform**, and **CB-RS**. Beneath this visual is the caption: **Figure 3: Visualization of features with Grad-CAM [17] on CIFAR100-LT. Uniform sampling mainly learns label-relevant features, while re-sampling overfits the label-irrelevant features.** The final section of text states: **On CIFAR100-LT, class-balanced re-sampling (CB-RS) leads to overfitting on the irrelevant contexts from tail classes, and unexpectedly affects the representation of head classes.** The footer indicates the slide number **10 / 17** and includes a URL link.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/bf6044d8-6d98-4015-ad31-59fedb1c633d_input_0.jpg"], "target_image": "images/multi_image/bf6044d8-6d98-4015-ad31-59fedb1c633d_target.png", "width": 1920, "height": 1080}
+{"index": 17, "id": "2fbc7e55-785c-4594-ba9e-655505162477", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "At the top of the slide is the title \"Motivation,\" with the Gaoling School of Artificial Intelligence logo and Chinese text in the upper right corner. Below, mathematical equations are displayed: $ d(\\boldsymbol{X},\\boldsymbol{Y})=\\langle\\boldsymbol{T}^{*},\\boldsymbol{D}\\rangle $ and $ \\boldsymbol{T}^{*}=\\underset{\\boldsymbol{T}\\in\\boldsymbol{\\Phi}}{\\arg\\min}\\langle\\boldsymbol{T},\\boldsymbol{D}\\rangle+\\mathcal{R}(\\boldsymbol{T}) $. Two bullet points follow: \"Inferring the optimal alignment $ T^{*} $ : solve an optimization problem under pre-defined constraints\" and \" $ \\Phi $ is the feasible set of T (temporal constraints).\" To the right, sub-image 1 illustrates a process flow: two vertically aligned color-coded blocks labeled X and Y point toward a large olive-green rounded rectangle containing the word \"optimization.\" Below, a smaller grayscale matrix labeled $ T^{*} $ sits above the word \"Alignment.\" Arrows connect D (a grayscale matrix on the left) to X and Y, and another arrow leads from D through a blue circle marked \"Frobenius dot product\" to the label \"Distance.\" A purple arrow also connects $ T^{*} $ to the blue circle, completing the feedback loop. The slide uses a clean white background with black and colored text, and decorative line art subtly appears in the corners.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/2fbc7e55-785c-4594-ba9e-655505162477_input_0.jpg"], "target_image": "images/multi_image/2fbc7e55-785c-4594-ba9e-655505162477_target.png", "width": 1920, "height": 1080}
+{"index": 18, "id": "8595e178-04bb-4590-abb7-3a41c852d2e4", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, titled **Problem Statement** in a large, dark grey sans-serif font at the top left, is set against a clean white background. In the upper left corner sits a small black logo resembling a stylized bird or abstract wing. The slide is divided into two primary sections: a text-heavy left column outlining the problem definition and a right column containing probabilistic graphical models. A thin, light grey horizontal line separates the title from the content below.\n\nThe left column features four bullet points, each marked by a small blue square. The first point states, **In this DA problem, the source data distribution $p_{S}(x,y)$ is different from the target data distribution $p_{T}(x,y)$.** Immediately following is the second point: **In the representation learning framework, we consider a representation mapping $p(z|x)$.** The third bullet introduces the output classifier, reading: **From z, we learn an output classifier/regressor $\\hat{p}(y|z)$. The predictive distribution is $\\mathbb{E}_{p(z|x)}[\\hat{p}(y|z)]$**. The final bullet point defines the loss function: **The loss on the target domain is $l_{test} = \\mathbb{E}_{p_{T}(x,y)}[-\\log \\mathbb{E}_{p(z|x)}[\\hat{p}(y|z)]]$**.\n\nOn the right side of the slide, there are two vertically stacked diagrams illustrating probability distributions. The top diagram is labeled **Source Distribution** in bold black text. It depicts a graph with three circular nodes labeled **y**, **x**, and **z**. A vertical arrow points from **y** down to **x**, and a horizontal arrow points from **x** to **z**. Beneath this diagram, the joint probability formula is written as: **Joint: $p_{S}(x,y,z) = p_{S}(x,y)p(z|x)$**. Directly below this is the second diagram, labeled **Target Distribution**. It mirrors the structure of the first, with a vertical arrow from **y** to **x** and a horizontal arrow from **x** to **z**. The accompanying text reads: **Joint: $p_{T}(x,y,z) = p_{T}(x,y)p(z|x)$**. At the very bottom right of the slide, the page number **4** is visible in small grey text. A dark blue footer bar runs across the entire bottom edge of the slide.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/8595e178-04bb-4590-abb7-3a41c852d2e4_input_0.jpg"], "target_image": "images/multi_image/8595e178-04bb-4590-abb7-3a41c852d2e4_target.png", "width": 2594, "height": 1460}
+{"index": 19, "id": "e335f966-119e-40f4-a4a4-d5b73e77029c", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide is titled \"Foundation models via contrastive learning\" in large, bold text at the top against a white background with a subtle gradient design element in the upper-right corner. The central visual component is a diagram, referred to here as sub-image 1, positioned in the middle of the slide. This sub-image 1 illustrates a flow where a central circle containing \"x\" branches out to two nodes labeled \"view 1\" and \"view 2\", passing through arrows marked \"t ~ T\" and \"t' ~ T\" respectively. The flow continues upward through circles labeled \"x_tilde_i\" and \"x_tilde_j\", then through functions \"f(.)\" to final outputs \"h_i\" and \"h_j\". Between these outputs, purple arrows point towards the center under the label \"<- Representation ->\". At the bottom-left corner of the slide, a logo and text read \"Stanford NLP | Alex Tamkin\", while the bottom-right corner displays the citation \"Chen et al, 2020\" and the slide number \"2\".\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/e335f966-119e-40f4-a4a4-d5b73e77029c_input_0.jpg"], "target_image": "images/multi_image/e335f966-119e-40f4-a4a4-d5b73e77029c_target.png", "width": 2508, "height": 1410}
+{"index": 20, "id": "c344318b-f9a3-4b75-9511-ba0ba9356c84", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide features the main title at the top reading \"$(2 + \\epsilon)$-approximate algorithm in $\\log n$ iterations.\" The layout is split into two sections. On the left side, a light green text box with a dashed blue border lists two bullet points defining the methodology: \"Given a threshold k, repeatedly remove all the vertices of degree less than $(2 + \\epsilon)k$\" and \"The approximate coreness value of a vertex is the largest k for which it is not removed.\" To the right of this text, the sub-image 1 displays a network graph with green nodes forming a central clique and red nodes on the periphery, accompanied by a red label \"k=1\" positioned just to the left of the diagram.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/c344318b-f9a3-4b75-9511-ba0ba9356c84_input_0.jpg"], "target_image": "images/multi_image/c344318b-f9a3-4b75-9511-ba0ba9356c84_target.png", "width": 1920, "height": 1080}
+{"index": 21, "id": "065771f5-cff6-4365-a1a4-fcc073e93659", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide is titled \"**Method**\" at the top in large bold text. The layout features a list of three mathematical formulations on the left and two corresponding flowcharts on the right. The first bullet point reads \"**Original**:\" followed by the equation \"$$ p(\\mathsf{v}_{t},\\mathsf{w}_{t})=p(\\mathsf{w}_{t})p(\\mathsf{v}_{t}) $$\". The second bullet point introduces \"**Structured Prior (Proposed)**:\" with the formula \"$$ p(\\mathsf{v}_{t},\\mathsf{w}_{t})=p(\\mathsf{w}_{t})p(\\mathsf{v}_{t}|\\mathsf{w}_{t}) $$\". The third item is \"**Temporal Prior (Expensive)**:\" accompanied by \"$$ p(\\mathsf{v}_{t},\\mathsf{w}_{t})=\\prod p(\\mathsf{w}_{t}|\\mathsf{w}_{<t})p(\\mathsf{v}_{t}|\\mathsf{v}_{<t}) $$\". To the right of this text, sub-image 1 shows a diagram labeled \"**SSF(Agustsson et al)**\" depicting a process starting with \"**$$ \\mathsf{z}_{t} $$**\" and a \"**split**\" box leading to \"**$$ \\mathsf{w}_{t} $$**\" and \"**$$ \\mathsf{v}_{t} $$**\", eventually feeding into functions like \"**$$ h_{warp} $$**\" and ending at \"**$$ \\mathsf{x}_{t} $$**\". Adjacent to it, sub-image 2 displays a diagram titled \"**Proposed**\" which similarly begins with \"**$$ \\mathsf{z}_{t} $$**\" and \"**split**\", but highlights a red box around \"**$$ \\mathsf{w}_{t} $$**\" and \"**$$ \\mathsf{v}_{t} $$**\" before routing through \"**$$ h_{\\sigma} $$**\" and \"**$$ h_{\\mu} $$**\" to reach \"**$$ \\mathsf{x}_{t} $$**\". The slide uses a clean white background with black text and color-coded shapes (purple, blue, green, red) within the flowcharts to distinguish components.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/065771f5-cff6-4365-a1a4-fcc073e93659_input_0.jpg", "images/multi_image/065771f5-cff6-4365-a1a4-fcc073e93659_input_1.jpg"], "target_image": "images/multi_image/065771f5-cff6-4365-a1a4-fcc073e93659_target.png", "width": 3840, "height": 2160}
+{"index": 22, "id": "54bc2aba-846b-48c5-af5b-9cf747ef9e48", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, titled \"Decision Region Quantification (DRQ)\" in a large, dark blue sans-serif font at the top left, outlines a methodology for improving predictive models. Below the title, the word \"Method\" appears in a smaller, lighter gray font. A thin, dark gray horizontal line separates the header from the main content. In the top right corner, three logos are displayed side-by-side: the Mila logo (purple hexagons), the Machine Learning Data Analytics logo (blue silhouette of a person with a cityscape), and the FAU logo (blue stylized letters). The main heading of the slide reads \"Quantify Robustness of Decision to improve predictions\" in a bold, black serif font. Below this heading, on the left side of the slide, is a vertical list of three steps labeled \"(a) Calibrate search radius\", \"(b) Explore candidate predictions\", and \"(c) Quantify robustness of candidates\", all in a standard black sans-serif font. To the right of this text is a 2x2 grid of four diagrams. The top-left diagram, labeled \"(a) Calibration\" beneath it, shows a data distribution map with a legend indicating \"Class 1\" (orange circles) and \"Class 2\" (blue circles). It depicts a dashed circular search radius centered on an orange point, crossing a white decision boundary into a blue region. The top-right diagram, labeled \"(b) Exploration\", is a variation of the first, showing the search radius with an additional blue star symbol near the boundary. The bottom-left diagram, labeled \"(c) Quantification\" and enclosed in a thin blue border, displays two separate dashed search circles, one in each class region, each containing a representative data point and a connecting arrow to a square marker. The bottom-right diagram, labeled \"(d) Decision boundary of the DRQ classifier\", shows the final result with the dashed circles removed, revealing a clean separation between the blue region on the left and the orange region on the right, with a single inverted triangle marker placed near the new decision boundary. At the bottom left of the slide, the text \"Technische Fakultät\" is visible in a small gray font. At the bottom right, the date \"21. Juni 2022\" and the page number \"7\" are printed in small black text.\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/54bc2aba-846b-48c5-af5b-9cf747ef9e48_input_0.jpg", "images/multi_image/54bc2aba-846b-48c5-af5b-9cf747ef9e48_input_1.jpg", "images/multi_image/54bc2aba-846b-48c5-af5b-9cf747ef9e48_input_2.jpg", "images/multi_image/54bc2aba-846b-48c5-af5b-9cf747ef9e48_input_3.jpg"], "target_image": "images/multi_image/54bc2aba-846b-48c5-af5b-9cf747ef9e48_target.png", "width": 1920, "height": 1080}
+{"index": 23, "id": "8da59bf4-3d19-40d6-9f18-792a2f6daae9", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide is titled \"Vision Encoder!\" in large, dark blue font at the top. Below the title, a caption reads: \"Table 8: Frozen encoder ablation. LiT style fine-tuning ablations on SugarCrepe, image-text retrieval, and ImageNet1k. Bold number indicates the best performance.\" This text introduces a comparative table that occupies the central portion of the slide, detailed further in sub-image 1. The table displays performance metrics across columns labeled \"Models,\" \"Train Text,\" \"Train Vision,\" \"SugarCrepe,\" \"Retrieval,\" and \"ImageNet1k.\" It compares results for \"LaCLIP\" and \"TripletCLIP (ours)\" under different training conditions. At the bottom of the slide, two bullet points summarize the findings, stating: \"Vision encoder is also important for the compositionality.\" followed by: \"So, it is important for future works to consider this modality!\"\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/8da59bf4-3d19-40d6-9f18-792a2f6daae9_input_0.jpg"], "target_image": "images/multi_image/8da59bf4-3d19-40d6-9f18-792a2f6daae9_target.png", "width": 1920, "height": 1080}
+{"index": 24, "id": "4be29d0d-5dbe-49b1-b92b-fb796e32b424", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The slide opens with the title **Two alternatives: “precommitted” and “nested” CVaR** at the top. Below this, on the left side, is the section heading **“Fixed” CVaR**, followed by the explanatory text: **CVaR, with same alpha, is applied at every stage to the outcome distribution**. To the right of this text, **sub-image 1** is positioned in the upper-right area, displaying a diagram labeled **start** which splits into branches marked **p(b|s) = 0.1** and **p(g|s) = 0.9**, leading to distributions labeled **bad state** and **good state**, all annotated with **CVaR<sub>α=0.1</sub>**.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/4be29d0d-5dbe-49b1-b92b-fb796e32b424_input_0.jpg"], "target_image": "images/multi_image/4be29d0d-5dbe-49b1-b92b-fb796e32b424_target.png", "width": 2642, "height": 1475}
+{"index": 25, "id": "4f06a40f-b848-4f16-a76b-874d7bfb8f31", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide from ShanghaiTech University is titled \"How to Construct Robust Pseudo-Labeling?\" in a prominent red banner at the top. The layout contrasts two methodologies, beginning with a section next to sub-image 1 (a sad face icon) labeled \"Existing Prediction-based PL,\" which lists the limitations: it \"evaluate each sample independently\" and \"heavily rely on the unreliable model's prediction.\" Below this, a section next to sub-image 2 (a smiling face icon) introduces \"Optimal Transport (OT) -based PL,\" noting that it further \"consider the inter-distribution structure of the samples and categories distribution\" and \"produce pseudo labels with global discriminability.\" The bottom half features two scatter plots side-by-side: the left plot is titled \"Prediction-based PL\" and the right is titled \"OT-based PL.\" Next to the right plot, the text \"global discriminability\" appears with a checkmark, visually comparing the two approaches against a white background featuring red architectural silhouettes at the bottom.\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/4f06a40f-b848-4f16-a76b-874d7bfb8f31_input_0.jpg", "images/multi_image/4f06a40f-b848-4f16-a76b-874d7bfb8f31_input_1.jpg", "images/multi_image/4f06a40f-b848-4f16-a76b-874d7bfb8f31_input_2.jpg", "images/multi_image/4f06a40f-b848-4f16-a76b-874d7bfb8f31_input_3.jpg"], "target_image": "images/multi_image/4f06a40f-b848-4f16-a76b-874d7bfb8f31_target.png", "width": 1920, "height": 1080}
+{"index": 26, "id": "ba8cddb0-6139-4377-922a-f66f76808861", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The slide features the title \"X Becomes a proper standard\" in bold purple font at the top left, accompanied by a large graphic on the right showing a black X with an orange elliptical ring, identified as sub-image 1. A bulleted list below the title details the standardization process, starting with \"Eventually Motif would become the default widget set\" followed by \"GTK, Qt and others would be built off of Xlib.\" The third point reads \"Extensions would expand the protocol with optional features,\" while the next line states \"After sometime XFree86 would become de facto X provider.\" The list continues with \"Eventually this would all collapse into a single group the Xorg Foundation taking over maintenance of X\" and concludes with \"X is ported to Linux.\" The footer includes \"Integrated Computer Solutions Inc.\" and the website \"www.ics.com\" alongside an ICS logo.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/ba8cddb0-6139-4377-922a-f66f76808861_input_0.jpg"], "target_image": "images/multi_image/ba8cddb0-6139-4377-922a-f66f76808861_target.jpg", "width": 2048, "height": 1152}
+{"index": 27, "id": "1beca56d-b01e-4da8-b366-89248c661a1c", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "Set against a solid, deep navy-blue background, this presentation slide utilizes a structured layout with three distinct yellow circles aligned horizontally along the top edge, serving as a decorative header element. Centered near the top, directly beneath the yellow circles, is a rectangular image (sub-image 1) that occupies roughly the upper center portion of the visual space. The central focus of the slide is a warning message displayed in a bold, italicized sans-serif font. The text reads: \"Ignoring fuel injector problems in your Porsche can compromise power and engine health.\" This message is split across the center; the phrase \"Ignoring fuel injector\" and \"problems in your Porsche can\" are colored in a bright yellow, creating a strong contrast against the blue, while the phrase \"compromise power and engine health\" is rendered in white.\n\nPositioned in the lower-left quadrant is a second rectangular image (sub-image 2), which sits adjacent to the text block. To the right of the main text body, slightly below the vertical center, is a third rectangular image (sub-image 3). These visual elements frame the textual warning. Anchoring the bottom center of the slide is the brand identifier \"BAY DIAGNOSTIC\" in a heavy, uppercase sans-serif font; the word \"BAY\" appears in white, while \"DIAGNOSTIC\" is styled in a bright cyan blue, distinguishing the company name at the foot of the composition. The overall design relies on a clean, high-contrast color palette of navy, yellow, white, and cyan to highlight the technical automotive content.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/1beca56d-b01e-4da8-b366-89248c661a1c_input_0.jpg", "images/multi_image/1beca56d-b01e-4da8-b366-89248c661a1c_input_1.jpg", "images/multi_image/1beca56d-b01e-4da8-b366-89248c661a1c_input_2.jpg"], "target_image": "images/multi_image/1beca56d-b01e-4da8-b366-89248c661a1c_target.jpg", "width": 2048, "height": 1152}
+{"index": 28, "id": "cf5bae6f-f890-43b3-bb36-4a42797e8228", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "Titled \"Behavior of Plastics,\" this presentation slide displays a stress-strain chart containing sub-image 1 on the left, which illustrates curves for rigid, tough, and soft materials. To the right of the chart, the caption reads: \"Figure 7.8 General terminology describing the behavior of three types of plastics. PTFE (polytetrafluoroethylene) has Teflon as its trade name. Source: R. L. E. Brown.\" The slide features a textured background with a footer that lists \"Kalpakjian • Schmid Manufacturing Engineering and Technology,\" \"© 2001 Prentice-Hall,\" and \"Page 7-11.\"\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/cf5bae6f-f890-43b3-bb36-4a42797e8228_input_0.jpg"], "target_image": "images/multi_image/cf5bae6f-f890-43b3-bb36-4a42797e8228_target.jpg", "width": 2048, "height": 1582}
+{"index": 29, "id": "62de33e2-3854-4924-a6a1-d872656970b6", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "At the top of the slide, the phrase \"Routine - Reduce - Restrict - Remix\" is written in a handwritten-style font, with the final word \"Remix\" highlighted in red and underlined. Directly below, the text reads, \"I’m collecting thumbnails, titles, and diagrams I love. (And used them as inspiration for this slide deck)\". In the center of the layout, a dark gray rectangular banner displays the text \"# swipe_file\". The lower half features three images arranged horizontally: sub-image 1 is positioned on the left, sub-image 2 is in the center, and sub-image 3 is on the right, which is a YouTube thumbnail with a red background featuring a portrait and the text \"Norm Macdonald is a comic genius\". At the very bottom right corner, the handle \"@shaanvp\" is visible.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/62de33e2-3854-4924-a6a1-d872656970b6_input_0.jpg", "images/multi_image/62de33e2-3854-4924-a6a1-d872656970b6_input_1.jpg", "images/multi_image/62de33e2-3854-4924-a6a1-d872656970b6_input_2.jpg"], "target_image": "images/multi_image/62de33e2-3854-4924-a6a1-d872656970b6_target.jpg", "width": 2048, "height": 1153}
+{"index": 30, "id": "c4daee21-6aaf-4dd0-a06d-7573e02ff619", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, professional layout with a dominant white background and a bold, solid orange vertical bar running along the entire left edge. Inside this orange bar, the word \"Assumptions\" is printed vertically in large, white sans-serif letters, oriented to be read from bottom to top. In the top right corner of the white area, a multi-colored logo for \"civil civic alonspace\" is visible, accompanied by Greek text reading \"Χώρος Πολιτών\". Just below the logo area, centered horizontally, is the slide title: \"- Relationship between assumptions and objective hierarchy\", written in black text.\n\nThe central focus of the slide is a diagram, identified in the requirements as a sub-image, which occupies the middle portion of the frame. This diagram illustrates a vertical hierarchy on the left side, consisting of four light blue rectangular boxes stacked vertically. From top to bottom, these boxes are labeled \"Overall Objective\", \"Purpose\", \"Results\", and \"Activities\". Below the \"Activities\" box, a black arrow points upward towards it, labeled \"Inputs\". To the right of this stack, three separate light blue boxes labeled \"Assumptions\" are positioned at the same vertical levels as \"Activities\", \"Results\", and \"Purpose\" respectively. Blue dashed arrows connect these elements: one points from the lowest \"Activities\" box to the lowest \"Assumptions\" box, another from \"Results\" to its corresponding \"Assumptions\" box, and a third from \"Purpose\" to its \"Assumptions\" box. Small plus signs (+) appear on these horizontal arrows. Additionally, diagonal dashed arrows point upward and to the left, connecting the \"Assumptions\" boxes back towards the hierarchy above them (e.g., from the bottom \"Assumptions\" box up to the \"Results\" box).\n\nTextual annotations clarify the logic of this diagram. Situated in the upper right area of the diagram, italicized text reads: \"e.g if activities are undertaken AND assumptions hold true, then results can be achieved, etc\". In the lower right area, below the bottom-most assumption connection, another italicized note explains: \"Pre-conditions – need to be met before resources are committed and activities initiated\".\n\nBelow the diagram area, centered on the slide, is a concluding paragraph of text in standard black font. This text reads: \"The assumptions are part of the vertical logic in the LogFrame. Start from the bottom of the matrix and work upwards.\" The overall visual style is minimalistic, utilizing a color palette of white, orange, light blue, and black to clearly distinguish structural elements and information flows.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/c4daee21-6aaf-4dd0-a06d-7573e02ff619_input_0.jpg"], "target_image": "images/multi_image/c4daee21-6aaf-4dd0-a06d-7573e02ff619_target.jpg", "width": 2048, "height": 1536}
+{"index": 31, "id": "4ee03cb8-149c-48f2-981e-fcefbb994a81", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This slide is titled \"Functions of Carbohydrates,\" which appears in bold, large text in the upper-left section. Below the title is an orange divider line, followed by a list of two functions, each marked with a checkmark. The first bullet point reads \"Structural support,\" and the second reads \"Storage of energy.\" These are located on the left side of the slide against a white background. The right side features a large graphical element, sub-image 1, which depicts a stylized blue molecular structure with floating spheres, set against a blue-toned, microscopic background.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/4ee03cb8-149c-48f2-981e-fcefbb994a81_input_0.jpg"], "target_image": "images/multi_image/4ee03cb8-149c-48f2-981e-fcefbb994a81_target.jpg", "width": 2048, "height": 1152}
+{"index": 32, "id": "43b748cb-d2a7-4615-b7b0-58802fe28431", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, set against a clean white background with a prominent dark purple border along the top and bottom edges, introduces an advanced artificial intelligence model. The slide is titled in large, dark blue text at the top: \"BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi- -Granularity Text Embeddings Through Self-Knowledge Distillation\". The layout is split into two main sections: a list of technical specifications on the left and a feature diagram on the right.\n\nOn the left side, the text identifies the model's origin with the line \"From Beijing Academy of Artificial Intelligence\". Immediately below, it states \"More than 100 languages supported\". A bulleted list follows, detailing specific architectural parameters. The first bullet reads \"- window: 8192 tokens!\". The second bullet specifies \"- parameters: 567M\". The third indicates \"- emb dim: 1024\". The fourth lists \"- pooling: [CLS] token\". The final bullet describes the underlying architecture: \"- pre-trained model: XLM-RoBERTa large (24 transformer's encoder layers)\". At the very bottom left corner, within the purple border, is the logo and text \"Sease | Information Retrieval Applied\".\n\nOn the right side of the slide, a visual breakdown illustrates the core capabilities of the model, organized into three columns under bold headers: \"Multi-Linguality\", \"Multi-Functionality\", and \"Multi-Granularity\". Under \"Multi-Linguality\", three rounded rectangular boxes are stacked vertically, reading \"100+ Languages\", \"Multi-Lingual\", and \"Cross-Lingual\" from top to bottom, all enclosed within a larger dashed border. The middle column, \"Multi-Functionality\", similarly contains three stacked boxes within a dashed outline, labeled \"Dense Retrieval\", \"Sparse Retrieval\", and \"Multi-Vec Retrieval\". The rightmost column, \"Multi-Granularity\", features three boxes labeled \"Sentence-Level\", \"Passage-Level\", and \"Doc-Level (≤8192)\". Beneath these three columns sits a single, wider, light gray box spanning the width of the diagram, labeled \"M3-Embedding\", indicating the unifying nature of the model. In the bottom right corner of the purple footer, the slide number \"28\" is displayed in white text.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/43b748cb-d2a7-4615-b7b0-58802fe28431_input_0.jpg"], "target_image": "images/multi_image/43b748cb-d2a7-4615-b7b0-58802fe28431_target.jpg", "width": 2048, "height": 1152}
+{"index": 33, "id": "7174cf86-ca8a-43d9-83bb-293b9e8fcec3", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a vibrant, split-screen design that balances a rich, illustrative graphic on the left with clean, informative text on the right. The overall aesthetic is modern and tech-focused, utilizing a palette of electric blues, warm oranges, and soft pinks.\n\nThe left half of the slide is dominated by sub-image 1, a detailed, colorful illustration that occupies the full vertical height of that side. This artwork depicts a stylized male figure with dark skin and short hair, wearing blue-rimmed glasses and a bright blue hoodie featuring a rainbow graphic on the chest. He sits cross-legged in a meditative yet active pose on a background filled with glowing circuit board lines and network nodes. On his lap rests an open laptop with the word \"CYBERSECURITY\" visible across the top bezel of the screen. The figure is holding a stylus or pen in his right hand, poised as if interacting with a digital interface floating just above the keyboard. The surrounding space is crowded with various digital icons connected by thin, colorful lines, including a ghost-like figure in the upper left, a blue cloud, a red location pin, an envelope symbol, gears, a padlock, and a magnifying glass with an eye, all rendered in a flat, vector-art style that suggests a complex digital ecosystem.\n\nTo the right of this illustration, the background shifts to a clean, white space that houses the slide's textual content. At the top of this section, the main title is displayed in a large, bold, black sans-serif font, reading \"Gen-AI for the Blue Team.\" Below the title, separated by a generous amount of white space, is a descriptive paragraph in a smaller, regular-weight, dark grey sans-serif font. This text provides the core message of the slide, stating verbatim: \"Practical guidance about how to match security needs with LLM strengths when evaluating potential AI solutions for cybersecurity defense.\" The text is left-aligned and spans three lines, creating a balanced visual block that contrasts sharply with the busy, colorful illustration on the left. There are no other graphical elements, borders, or charts present on the right side, ensuring the focus remains on the headline and the explanatory guidance provided.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/7174cf86-ca8a-43d9-83bb-293b9e8fcec3_input_0.jpg"], "target_image": "images/multi_image/7174cf86-ca8a-43d9-83bb-293b9e8fcec3_target.jpg", "width": 2048, "height": 1152}
+{"index": 34, "id": "daf9a97b-d727-4ea6-9f1d-69e8621c3e16", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide utilizes a clean, minimalist design with a stark white background to highlight technical content and a visual pun. In the upper-left corner, the title is displayed as a large, bold, dark grey heading reading **ACID**. Below this heading, a primary bullet point introduces the concept, stating: **• RDS databases offer full ACID compliance**. Indented beneath this main point is a list of four sub-bullets that spell out the acronym, listing **• Atomicity**, followed by **• Consistency**, **• Isolation**, and finally **• Durability**. All text in this left-hand column is rendered in a simple, sans-serif dark grey font.\n\nTo the right of the text, occupying the middle-right section of the slide, is sub-image 1, a rectangular photograph showing several glass laboratory bottles with black stoppers, arranged closely together. The labels on the bottles are handwritten in black ink; the central bottle is clearly marked **H2SO4 2N**, while a bottle to its left shows **NaOH**, and one to its right displays **HND3 2**, serving as a visual metaphor for the slide's topic.\n\nThe slide features a distinct footer area. Running vertically along the far right edge, rotated 90 degrees counter-clockwise, is a disclaimer in grey text that reads **NOT FOR DISTRIBUTION © Stephane Marek www.datacumulus.com**. The bottom of the slide is anchored by a bright orange graphical banner. In the lower-left corner, over the orange background, small white text lists **sundog-education.com** and **datacumulus.com**, followed by the copyright notice **© 2023-2026 All Rights Reserved Worldwide**. Opposite this, on the lower-right side, the orange banner contains two logos: the blue cloud logo for **DataCumulus** and the logo for **Sundog Education**, which includes a stylized arc. The overall aesthetic is professional and educational, utilizing the chemical imagery to reinforce the database terminology presented in the text.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/daf9a97b-d727-4ea6-9f1d-69e8621c3e16_input_0.jpg"], "target_image": "images/multi_image/daf9a97b-d727-4ea6-9f1d-69e8621c3e16_target.jpg", "width": 2048, "height": 1152}
+{"index": 35, "id": "0cdf2648-51b9-49dd-a26a-2f1534d39b1c", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide features a light beige background with a red banner-like arrow in the upper-left corner, introducing the topic under the header **CBC (Cake Breaker Conveyor)** framed in a rounded rectangle. On the left side, the section **CBC 1** lists **Transmisi : Poros langsung** and **Gear motor:**. Below that, **CBC 2** specifies **Transmisi : sprocket - rantai**. This area includes a diagram showing **Gear Motor** and **Conveyor** connected by a belt, where the motor side has a diameter of **30 cm**, the conveyor side is **20 cm**, and the distance between them is **60 cm**. This diagram corresponds to **sub-image 1** located in the lower-left area. Beneath this diagram, the text notes **Rantai : 1.5 inch** and **Gear motor :**. On the right side, the section **CBC 3** indicates **Transmisi : Pully-Belt**. It displays a schematic labeled **Elektromotor** on the left and **Gearbox** on the right, illustrating a belt drive with a motor pulley of **37 cm**, a gearbox pulley of **26 cm**, and a center distance of **66 cm**. This visual corresponds to **sub-image 2** positioned in the upper-right area. Below the diagram for CBC 3, the specifications are listed as **Gear motor : 15 Kw, 1460 rpm** and **Jenis V-belt: C**.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/0cdf2648-51b9-49dd-a26a-2f1534d39b1c_input_0.jpg", "images/multi_image/0cdf2648-51b9-49dd-a26a-2f1534d39b1c_input_1.jpg"], "target_image": "images/multi_image/0cdf2648-51b9-49dd-a26a-2f1534d39b1c_target.jpg", "width": 2048, "height": 1152}
+{"index": 36, "id": "839582bd-196e-44e9-881e-ce6983ecbbab", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Data Analysis\",\n    \"background\": {\n      \"color\": \"#F0F4F8\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#1E56A0\",\n      \"secondary\": \"#BDD7E8\",\n      \"text\": \"#000000\",\n      \"accent\": \"#1E56A0\"\n    },\n    \"special_instructions\": \"The slide features a clean, professional aesthetic typical of business presentations. The background is a very light, cool grey-blue. The design relies on high contrast between the dark blue title and the black body text against the light background. The chart uses a standard matplotlib style with a thin black border and a light blue fill for the confidence interval.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical\",\n    \"composition_style\": \"Asymmetric, Two-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 25,\n      \"blocks\": [\n        {\n          \"id\": \"title-25\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Launch Success Yearly Trend\",\n          \"style\": {\n            \"color\": \"#1E56A0\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold text aligned to the left. A thin horizontal line separates the title from the body content.\"\n          }\n        },\n        {\n          \"id\": \"separator-25\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A thin, solid horizontal line spanning the width of the slide, positioned directly below the title.\",\n          \"style\": {\n            \"illustration_style\": \"line\",\n            \"color_scheme\": \"Primary Blue (#1E56A0)\",\n            \"position\": \"below title, spanning full width\",\n            \"special_instructions\": \"Thin stroke width, solid color matching the title.\"\n          }\n        },\n        {\n          \"id\": \"body-25-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• From the plot, we can observe that success rate since 2013 kept on increasing till 2020.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text, black color, standard weight. Aligned to the left margin.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_791_323_1844_930.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_791_323_1844_930.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing launch success trend over years\",\n            \"caption\": \"Plot of launch success yearly trend\"\n          },\n          \"style\": {\n            \"position\": \"middle-right\",\n            \"image_style\": \"Standard plot frame with black border\",\n            \"container\": null,\n            \"special_instructions\": \"The chart displays a line plot with a shaded area representing a confidence interval. The line color is a medium blue, and the shaded area is a lighter blue. The x-axis is labeled 'Extracted_year' and the y-axis is labeled 'Class'.\"\n          }\n        },\n        {\n          \"id\": \"caption-25-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Plot of launch success yearly trend\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"inside chart area, top-center\",\n            \"text_style\": null,\n            \"special_instructions\": \"This text appears as the title inside the chart frame itself, not below it.\"\n          }\n        },\n        {\n          \"id\": \"page-number-25\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"25\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1E56A0\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small page number, colored in the primary blue theme color.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/839582bd-196e-44e9-881e-ce6983ecbbab_input_0.jpg"], "target_image": "images/multi_image/839582bd-196e-44e9-881e-ce6983ecbbab_target.jpg", "width": 2048, "height": 1152}
+{"index": 37, "id": "84559d22-5827-4a44-8092-90c1bde6e2dc", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, minimalist design with a stark white background and a heavy black footer bar running across the entire bottom edge. The slide's primary heading, located in the top-left corner in a large, bold sans-serif typeface, reads **\"Content Design: Best practices\"**, immediately followed by a secondary subtitle in a smaller, lighter font: **\"Contextual and Concise\"**. To the left of the slide's vertical centerline, a large circular graphic—referred to as sub-image 1—occupies a significant portion of the layout. On the right side, aligned to the right of the sub-image, is the main content area headed by the bold text **\"2. Create concise and contextual content\"**. Below this heading, a bulleted list provides specific guidelines. The first bullet point reads: **\"Always keep in mind that the content for Mobile has a different context, the approach needs to be tailored for this particular medium.\"** Nested beneath or following this are three additional points with key terms emphasized in bold: **\"Describe only what the user needs to know.\"**, **\"Eliminate redundancy, such as titles that restate the body of an information box.\"**, and **\"Keep text as short as possible.\"** In the bottom-right corner, sitting just above the black footer, is a small, vertical illustration of a white smartphone. The screen of this phone graphic displays the text **\"Mobile Best Practices\"** stacked in three lines. The overall aesthetic is professional and high-contrast, utilizing only black text on white, with the exception of the photographic colors within the sub-image.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/84559d22-5827-4a44-8092-90c1bde6e2dc_input_0.jpg"], "target_image": "images/multi_image/84559d22-5827-4a44-8092-90c1bde6e2dc_target.jpg", "width": 2048, "height": 1536}
+{"index": 38, "id": "32bccc69-7f3e-4ea8-a2e6-313afd9c11e1", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Technical Educational\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#009944\",\n      \"secondary\": \"#990000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#0000FF\"\n    },\n    \"special_instructions\": \"Clean, high-contrast technical slide. The background is pure white. The title uses a distinct green color. Key emphasis words in the body text are colored red. The diagram uses standard black lines with a specific blue accent for the positive feedback symbol.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Educational, Technical, Clear\",\n    \"composition_style\": \"Centered Title, Left-Aligned Body, Centered Diagram\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 11,\n      \"blocks\": [\n        {\n          \"id\": \"title-11\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Positive Feedback\",\n          \"style\": {\n            \"color\": \"#009944\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, green text centered at the top.\"\n          }\n        },\n        {\n          \"id\": \"body-11-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Positive feedback is the process in which the portion of output is added to the input, amplified again, and this process continues.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point. The word 'added' is highlighted in red (#990000).\"\n          }\n        },\n        {\n          \"id\": \"illustration-11-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A schematic block diagram of a positive feedback loop. It features an amplifier symbol (triangle labeled 'A') in the center. An 'input' line enters from the left. The output line exits to the right and is labeled 'output'. A feedback loop connects the output back to the input. The feedback path goes through a rectangular block labeled with the Greek letter beta (β). At the input junction, a blue circle with a white plus sign (+) indicates positive feedback.\",\n          \"style\": {\n            \"illustration_style\": \"schematic diagram\",\n            \"color_scheme\": \"black lines, blue accent circle\",\n            \"position\": \"center of the slide, between the two text blocks\",\n            \"special_instructions\": \"The feedback block is a simple rectangle. The amplifier is a standard triangle. The feedback symbol is a solid blue circle with a white plus sign inside.\"\n          }\n        },\n        {\n          \"id\": \"body-11-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Positive feedback is used in the design of oscillators and other applications.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"below the diagram, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point. Standard black text.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/32bccc69-7f3e-4ea8-a2e6-313afd9c11e1_input_0.jpg"], "target_image": "images/multi_image/32bccc69-7f3e-4ea8-a2e6-313afd9c11e1_target.jpg", "width": 2048, "height": 1536}
+{"index": 39, "id": "db6705c7-3272-4919-8c10-d12b4ee5154d", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide features a dark blue banner at the top displaying the text \"BIDS AND AWARDS COMMITTEE\". Below this header, the section \"BAC Honoraria\" appears in blue text, followed by the subheading \"Eligibility and Amount\". The first bullet point states that \"All BAC, TWG, and Ad Hoc Secretariat Members may receive honoraria up to 30% of their basic monthly salary, subject to fund availability.\" To the right of this text, sub-image 1 depicts two illustrated figures carrying stacks of coins. Continuing down the left column, a second bullet point reads: \"Undersecretaries, Assistant Secretaries, and equivalent ranks designated as BAC members are also entitled to honoraria.\" Further down, under the bold heading \"Exclusion,\" the text notes that \"Officials and employees whose primary function is procurement are not entitled to honoraria.\" At the bottom right, the citation \"(Reference: RA No. 12009, IRR Section 45)\" is displayed.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/db6705c7-3272-4919-8c10-d12b4ee5154d_input_0.jpg"], "target_image": "images/multi_image/db6705c7-3272-4919-8c10-d12b4ee5154d_target.jpg", "width": 2048, "height": 1152}
+{"index": 40, "id": "67c25ed3-9df9-4171-b006-34e11e157153", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean white background with a structured layout designed to provide a financial and operational overview of Arrowhead Pharmaceuticals Inc. At the top, a large, bold black title reads \"Arrowhead Pharmaceuticals Inc. Overview,\" underlined by a thin dark blue line. Immediately below the title on the left is a rectangular text box with a black border containing the description: \"Arrowhead Pharmaceuticals, Inc., headquartered in Pasadena, California, develops medicines for the treatment of intractable diseases by silencing the genes that cause them.\"\n\nThe main body of the slide is divided into two columns. The left column features two data tables stacked vertically. The top table, labeled \"sub-image 1\" in the input description but clearly visible here, compares statistics for two specific dates. The header row lists \"Statistic,\" \"6/8/15,\" and \"6/8/20.\" The rows below detail the following data points: \"Stock Price\" was $6.52 in 2015 and $35.74 in 2020; \"Market Cap\" grew from $387.93M to $3.64B; \"Enterprise Value\" increased from $272.23M to $3.34B; and \"Shares Outstanding\" rose from 59.50M to 101.77M. The valuation metrics show a significant change, with \"EV / NTM Revenue\" listed as \"N/A\" for 6/8/15 and \"32.54x\" for 6/8/20, while \"EV / NTM EBITDA\" and \"NTM P/E\" remain \"N/A\" for both dates. The table utilizes alternating light gray and white row backgrounds with dark blue headers.\n\nDirectly below this is a second table comparing fiscal years. The headers are \"Statistic,\" \"FY 2015,\" and \"FY 2019.\" The data reveals that \"Revenue\" for FY 2015 was $400K, whereas for FY 2019 it was $168.8M. Similarly, \"EBITDA\" improved from -$81.6M in FY 2015 to $65.6M in FY 2019. This table shares the same alternating row style and dark blue header design.\n\nOn the right side of the slide, the top section is titled \"NTM EV/Revenue Multiple.\" Below this title is a horizontal bar chart comparing two years. A yellow bar labeled \"2020\" indicates a value of \"32.5x.\" Beneath it, a much longer blue bar labeled \"2016\" extends significantly further to the right, indicating a value of \"287.1x.\" The horizontal axis is marked in increments of 100.0x, ranging from 0.0x to 400.0x.\n\nBelow the bar chart is a larger line graph titled \"Arrowhead Pharmaceuticals, Inc. (Nasdaq:GSRWR) - Share Pricing (Open: 6.52 High: 73.01 Low: 1.24 Close: 35.74 Avg: 13.81)\" and \"Arrowhead Pharmaceuticals, Inc. (Nasdaq:GSRWR) - TEV/Forward EBITDA (Open: 298.74 High: 298.74 Low: 160.40 Close: 160.40 Avg: 256.67).\" The graph plots data from September 2015 through May 2020 along the x-axis. It features a dual-axis system: the left y-axis represents share pricing ranging from 0 to 80, while the right y-axis represents multiples ranging from 150.00x to 325.00x. A cyan line tracks the share price, showing relative flatness until late 2018, followed by a steep surge to a peak near January 2020, and a subsequent sharp drop before a slight recovery. A thin brown line overlays parts of this data, tracking the TEV/Forward EBITDA multiple.\n\nThe footer of the slide contains a dark blue banner. On the bottom left sits the logo for \"ALTA FOX CAPITAL,\" and on the bottom right is the page number \"485\" in white text.\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/67c25ed3-9df9-4171-b006-34e11e157153_input_0.jpg", "images/multi_image/67c25ed3-9df9-4171-b006-34e11e157153_input_1.jpg", "images/multi_image/67c25ed3-9df9-4171-b006-34e11e157153_input_2.jpg", "images/multi_image/67c25ed3-9df9-4171-b006-34e11e157153_input_3.jpg"], "target_image": "images/multi_image/67c25ed3-9df9-4171-b006-34e11e157153_target.jpg", "width": 2048, "height": 1152}
+{"index": 41, "id": "5db9437a-8536-477a-9705-728849bb24f3", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide features a bold, high-contrast design with a minimalist background and a prominent photographic element. A sub-image, identified as sub-image 1, is positioned in the lower-left area, showing a white textured object with a black \"X\" mark. Above this, a dark curved shape occupies the upper-left quadrant against a stark white background.\n\nOn the right side of the slide, the text begins with the main headline: \"Never look at index in isolation and draw conclusion\". This title is underlined by a thick, hand-drawn orange line. Below this divider, a list of instructional points is presented in black text. The first bullet asks, \"How to do it?\". The subsequent points follow: \"Check advance decline\", then \"Find out Advance Decline ratio\", and finally concludes with the advice, \"The higher the better (for bullish stance)\".\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/5db9437a-8536-477a-9705-728849bb24f3_input_0.jpg"], "target_image": "images/multi_image/5db9437a-8536-477a-9705-728849bb24f3_target.jpg", "width": 2048, "height": 1152}
+{"index": 42, "id": "47894ac7-cde4-4fb0-a991-738485a04dd7", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This educational presentation slide features a deep purple-to-black gradient background, creating a high-contrast setting for the white and colored instructional elements. At the very top left, a rounded rectangular banner in a muted magenta shade contains the white text \"QUESTION (CBSE 2023)\". To the far right of this header, a white logo inside a shield-like shape displays the letters \"PW\", representing the brand. Below the banner, the main problem statement is presented in a bright yellow sans-serif font: \"The magnetic field lines near a substance are as shown in the figure. The substance is:\". Aligned to the far right of this sentence, the point value \"[1 Mark]\" is indicated in white brackets.\n\nThe left portion of the slide lists four multiple-choice options vertically, each preceded by a circular icon containing the option letter. Option A, labeled \"Copper\", is accompanied by a green checkmark icon to its left, indicating it as the correct answer. Option B, labeled \"Iron\", has a green circle drawn next to it containing the handwritten abbreviation \"F\". Option C, labeled \"Sodium\", features a green circle with the handwritten abbreviation \"Pa\" inside. Option D, labeled \"Aluminium\", similarly has a green circle with the handwritten abbreviation \"Pa\" inscribed within it.\n\nThe right half of the slide is dominated by a diagram illustrating magnetic field behavior, which acts as a sub-image referenced as sub-image 1 occupying the center-right area. Above this diagram, the word \"Diamagnetic\" is written in a handwritten, bright green font with a green underline emphasizing the term. The diagram itself depicts a hollow white rectangle centered horizontally. Surrounding this rectangle are several white curved lines with small arrowheads, representing magnetic field lines moving from left to right. These lines are noticeably pushed away from the central rectangle, bending outward to avoid entering the material, while maintaining their left-to-right flow direction. This visual arrangement demonstrates the repulsion of magnetic field lines by the substance inside the rectangle. The overall layout separates the textual question and options on the left from the explanatory diagram on the right, using the dark background to make the white lines and yellow text stand out clearly.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/47894ac7-cde4-4fb0-a991-738485a04dd7_input_0.jpg"], "target_image": "images/multi_image/47894ac7-cde4-4fb0-a991-738485a04dd7_target.jpg", "width": 2048, "height": 1152}
+{"index": 43, "id": "aa4f2e2c-fb64-4ce5-a3fe-8cd6c8347c9c", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean white background with a layout divided between textual bullet points on the left, a photograph on the right, and prominent warning text at the bottom. Centered at the very top in a large, black sans-serif font is the title \"Cable Installation - OSP Buried\". Beneath the title on the left side of the slide, a vertical list of five bullet points provides technical instructions in black text. The points read, in order: \"Know the install method\", \"Know limits in tension and bend radius\", \"Attach pulling eyes properly\", \"Despool by rolling off the spool\", and \"Figure 8 for midspan pulls\".\n\nTo the right of this text list, occupying the upper-right quadrant of the slide, is sub-image 1. This rectangular image depicts an outdoor industrial scene under a bright sky, showing workers operating heavy machinery and large cable spools on a dirt field. At the bottom of the slide, spanning the width of the content area, is a bold, capitalized directive in black text that reads \"CALL BEFORE YOU DIG: Dial 811\".\n\nBelow this warning message, centered at the very bottom, is a line of smaller gray text serving as a copyright notice: \"CFOT Curriculum © 2022, The Fiber Optic Assn. Inc.\" In the bottom-left corner, there is sub-image 2, a small circular logo containing the letters \"FOA\" in blue against a white background with a blue border. In the bottom-right corner, the slide number \"232\" appears in small black text. The overall visual style is professional and instructional, utilizing high-contrast black text on white to ensure readability of the technical guidelines and safety warnings.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/aa4f2e2c-fb64-4ce5-a3fe-8cd6c8347c9c_input_0.jpg", "images/multi_image/aa4f2e2c-fb64-4ce5-a3fe-8cd6c8347c9c_input_1.jpg"], "target_image": "images/multi_image/aa4f2e2c-fb64-4ce5-a3fe-8cd6c8347c9c_target.jpg", "width": 2048, "height": 1536}
+{"index": 44, "id": "05145538-20f8-49f0-bbc6-71c3048b7dcb", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "Set against a light green background, this presentation slide features a darker green header banner reading \"WHAT IS GREEN TECHNOLOGY\" in white text, with the number \"3\" visible in a circle on the top left. The main content is arranged in three horizontal rows, each containing a visual on the left and text on the right. The first row displays sub-image 1 next to the statement: \"“Green Technology” is a system that uses innovative methods to create an environmental friendly products\". Below this, the second row presents sub-image 2 alongside the text: \"It uses renewable natural resources that never depletes, so future generation can also benefit from it.\" The third row shows sub-image 3, next to which the label \"BCM 501\" appears above the sentence: \"It can effectively change waste pattern and production in a way that it won't harm the planet.\" Finally, the date \"24/11/2014\" is printed in the bottom right corner.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/05145538-20f8-49f0-bbc6-71c3048b7dcb_input_0.jpg", "images/multi_image/05145538-20f8-49f0-bbc6-71c3048b7dcb_input_1.jpg", "images/multi_image/05145538-20f8-49f0-bbc6-71c3048b7dcb_input_2.jpg"], "target_image": "images/multi_image/05145538-20f8-49f0-bbc6-71c3048b7dcb_target.jpg", "width": 2048, "height": 1536}
+{"index": 45, "id": "dfeaa6fe-8991-4ac6-a87f-f3569f13411c", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "In the top left corner, a green Alfa logo (labeled 'A') sits above a paragraph of green text stating: \"To kick off this partnership, Alfa’s headquarters and flagship facility, Alfa Campus, has gone through the GNFZ certification process.\" The slide layout is split diagonally; the right side is dedicated to a sub-image showing a bright, modern office interior with wooden beams and glass walls. At the bottom left, the \"GLOBAL NETWORK for ZERO\" logo is displayed, with the same text repeated in the bottom right corner.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/dfeaa6fe-8991-4ac6-a87f-f3569f13411c_input_0.jpg"], "target_image": "images/multi_image/dfeaa6fe-8991-4ac6-a87f-f3569f13411c_target.jpg", "width": 2048, "height": 1152}
+{"index": 46, "id": "eb58d62e-3f10-4be9-ac83-71c996daac2a", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide highlights the culinary traditions of the region, featuring the title \"THE RICH HERITAGE OF INDONESIAN FOOD\" displayed in large, bold, orange letters inside a dark brown banner on the left. Below this header, a paragraph of white text against an orange background explains that \"The rich heritage of Indonesian food reflects centuries of diverse cultural influences, vibrant spices, and traditional cooking techniques that create a unique and flavorful culinary experience.\" The right side of the slide includes sub-image 1, positioned vertically, which shows close-up photography of various dishes including a meat stew in a white bowl and ingredients arranged on banana leaves.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/eb58d62e-3f10-4be9-ac83-71c996daac2a_input_0.jpg"], "target_image": "images/multi_image/eb58d62e-3f10-4be9-ac83-71c996daac2a_target.jpg", "width": 2048, "height": 1152}
+{"index": 47, "id": "73d640f4-c397-4b70-b9df-7b4cc8f85e50", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide from the \"World Energy Investment 2025\" report is titled \"Demand and electrification\" in an orange banner at the top right. The main headline states that \"Boosted by strong EV sales and efficiency gains in construction and industry, demand-side investment has nearly doubled since 2015.\" The visual centerpiece, labeled as sub-image 1, is positioned in the center of the slide and is titled \"Global investment in energy efficiency, electrification and renewables for end uses by sector 2015–2025.\" This stacked bar chart tracks investment in Billion USD (2024, MER) from 2015 through an estimated 2025, with a legend identifying sectors as \"Buildings,\" \"Transport,\" and \"Industry.\" Below the chart, the text explains that \"The levelling off of EV sales in some parts of the world was largely compensated by the increasing sales of Chinese vehicles in large EMDE markets, while investment in more efficient appliances kept investment in the buildings sector from declining.\" The footer contains a section labeled \"Notes\" which reads: \"EMDE = emerging market and developing economies; 2025e = estimated values for 2025. An energy efficiency investment is defined as the incremental spending on new energy-efficient equipment or the cost of refurbishments that reduce energy use (excluding labour). The intention is to capture additional spending compared to the reference technology and which leads to reduced energy consumption. The methodology for tracking demand investment was updated in 2025 and now tracks a broader range of technologies. Renewables for end use include behind-the-meter technologies such as solar water and geothermal heating.\" The page is marked \"PAGE | 83\" and features the IEA logo in the bottom right corner.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/73d640f4-c397-4b70-b9df-7b4cc8f85e50_input_0.jpg"], "target_image": "images/multi_image/73d640f4-c397-4b70-b9df-7b4cc8f85e50_target.jpg", "width": 2048, "height": 1448}
+{"index": 48, "id": "62310b50-4ea9-4527-8c85-64e6bf0f7069", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The slide presents a stacked bar chart titled \"IMP & TFWP Work Permit Holders—2018-25\" centered at the top in large, bold black font. The chart visualizes data trends across years, with the horizontal axis labeled from \"2018-Q1\" through \"2025-Q3\" and a vertical axis marked with values ranging from 100,000 to 400,000. A legend at the bottom identifies the colored segments as \"Agreements,\" \"Agriculture,\" \"Cdn Interests,\" \"LMIA,\" \"Other IMP,\" \"Vulnerable,\" \"Not Stated,\" and \"Caregivers,\" alongside a page number \"25\" at the very bottom center. The visual content of the chart is also referenced as sub-image 1.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/62310b50-4ea9-4527-8c85-64e6bf0f7069_input_0.jpg"], "target_image": "images/multi_image/62310b50-4ea9-4527-8c85-64e6bf0f7069_target.jpg", "width": 2048, "height": 1152}
+{"index": 49, "id": "2516a8ea-24da-4a28-b1ec-7e24bfe6bad8", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide, titled \"Local Factors\" in blue text, introduces the topic of \"Plaque and Calculus\" in green, addressing \"The microbial challenge presented by bacterial plaque and calculus is the most important local factor in periodontal diseases\" as the primary bullet point. The slide emphasizes that \"Therefore, in most cases, having a good prognosis depends on the ability of the patient and the clinician to remove these etiologic factors\" in a second bullet point. Visually, the layout features a white background with these textual explanations centered above two comparative images positioned side-by-side at the bottom: sub-image 1 is positioned on the left, and sub-image 2 is positioned on the right.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/2516a8ea-24da-4a28-b1ec-7e24bfe6bad8_input_0.jpg", "images/multi_image/2516a8ea-24da-4a28-b1ec-7e24bfe6bad8_input_1.jpg"], "target_image": "images/multi_image/2516a8ea-24da-4a28-b1ec-7e24bfe6bad8_target.jpg", "width": 2048, "height": 1536}
+{"index": 50, "id": "a71aecce-b47a-4208-be22-077a7d401631", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "Titled \"CULTURE METHODS\" in large yellow text against a blue header bar, this slide outlines specific \"Culture techniques.\" Below, an orange bullet point introduces the \"Streak plate Method or Surface plating,\" followed by a sub-bullet identifying the \"Quadrat streak.\" To the left, sub-image 1 displays a peach-colored circle divided into four quadrants labeled 1 through 4. To the right, sub-image 2 illustrates a petri dish diagram detailing the process: beginning at the bottom left with \"Apply loopful of culture (start)\" which results in \"Heavy confluent growth.\" Moving to the top left, labeled [1], an arrow points up labeled \"Apply flamed loop.\" Moving right into quadrant [2], the label indicates \"Heavy growth\" after another \"Apply flamed loop\" step. Further right into quadrant [3], an arrow points down labeled \"Apply flamed loop\" resulting in \"Light growth.\" Finally, in the bottom center labeled [4], a leftward arrow labeled \"Apply flamed loop\" points to the area showing \"Discrete colonies.\"\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/a71aecce-b47a-4208-be22-077a7d401631_input_0.jpg", "images/multi_image/a71aecce-b47a-4208-be22-077a7d401631_input_1.jpg"], "target_image": "images/multi_image/a71aecce-b47a-4208-be22-077a7d401631_target.jpg", "width": 2048, "height": 1536}
+{"index": 51, "id": "1c0ef511-e4c2-41c9-930f-4395e8ebad22", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide is titled **PROBLEM STATEMENT - 13**, positioned prominently at the top. The slide features a decorative background composed of geometric cubes in shades of orange and maroon, clustered on the left and right edges. Centered in the main body is a bullet point that reads: **• A descriptive study to identify the cultural beliefs regarding perinatal dietary habit among women.** The slide includes two sub-images: sub-image 1 is located in the lower-left area, and sub-image 2 is positioned in the upper-right area, both appearing as cropped sections of the geometric design.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/1c0ef511-e4c2-41c9-930f-4395e8ebad22_input_0.jpg", "images/multi_image/1c0ef511-e4c2-41c9-930f-4395e8ebad22_input_1.jpg"], "target_image": "images/multi_image/1c0ef511-e4c2-41c9-930f-4395e8ebad22_target.jpg", "width": 2048, "height": 1152}
+{"index": 52, "id": "4024df08-d9df-4571-b8d5-63aff181ae33", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, professional design set against a deep navy blue background that transitions subtly into a slightly lighter shade towards the bottom right. In the top-left corner, a white logo for \"sdle corp\" sits adjacent to the tagline \"Your Technology Partner.\" Dominating the upper center of the slide is the main title \"Booking & Reservation\" written in large, bold white sans-serif type. Below the title, on the left side of the slide, is a bulleted list containing three key value propositions. The first bullet point reads \"Streamline bookings with Odoo POS mobile app solutions.\" followed immediately by the second point, \"Simplify reservation management using Odoo POS development.\" The final bullet concludes the text section with \"Odoo POS ensures easy reservation handling for businesses.\" All text in this list is white, matching the title, and is aligned to the left.\n\nTo the right of the text, occupying roughly the right third of the slide, is a colorful illustration (sub-image 1) that visually represents the booking process. This illustration is centered around a large, vertical smartphone displaying a mobile application interface. On the phone screen, the header displays the word \"HOTEL\" above a graphic of a red building. Below the building icon, there are interface elements including a green phone icon, a row of five yellow stars indicating a rating, and a prominent orange button at the bottom labeled \"BOOK NOW.\" To the right of the smartphone stands a stylized cartoon man in a green shirt and black pants, holding a credit card in one hand and gesturing toward the phone with the other. Surrounding this central action are various travel and digital icons: a large orange map pin marker to the far right, a blue \"thumbs up\" speech bubble near the man's waist, a green chat bubble above his head, and a blue calendar with a date circled and crossed out sitting at the bottom left of the phone. Subtle green leaf elements are tucked behind the phone and calendar, adding a touch of organic texture to the graphic. A faint white diagonal line cuts across the bottom left corner of the slide, separating the dark background from a slightly lighter geometric shape below it.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/4024df08-d9df-4571-b8d5-63aff181ae33_input_0.jpg"], "target_image": "images/multi_image/4024df08-d9df-4571-b8d5-63aff181ae33_target.jpg", "width": 2048, "height": 1152}
+{"index": 53, "id": "1c2e57d6-51a5-4d53-ab45-fcfb72b720c9", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, set against a clean white background, is titled \"MPLS LFIB with Segment Routing\" in large, dark blue sans-serif text at the top left. Below the title, the slide is organized into three main bullet points on the left side, accompanied by visual aids on the right and bottom. The first point states, \"LFIB populated by IGP (ISIS / OSPF),\" followed by a smaller sub-line reading, \"More efficient label allocation than LDP.\" The second point highlights, \"Source base routing: no state in P nodes.\" The third point declares, \"Forwarding table remains constant,\" with a parenthetical explanation below it: \"(Nodes + Adjacencies) regardless of number of paths.\"\n\nVisually, the upper right quadrant features a network topology diagram (sub-image 2). This diagram illustrates a cloud-shaped area containing multiple router icons labeled \"PE\" and \"P,\" interconnected by a dense mesh of yellow lines representing paths.\n\nIn the lower left quadrant, a bar chart is displayed (sub-image 1). The chart is titled \"Number of states at a core node.\" Its vertical axis ranges from 0 to 1,000,000 in increments of 100,000, while the horizontal axis marks values of 50, 100, 300, 500, and 1000. The legend identifies blue bars as \"SR\" and red bars as \"Full-Mesh TE.\" The data shows that for every x-axis category, the red \"Full-Mesh TE\" bars grow significantly larger as the number increases, whereas the blue \"SR\" bars remain flat and near zero across all categories.\n\nIn the lower right quadrant, a forwarding table (sub-image 3) is presented with three columns: \"In Label,\" \"Out Label,\" and \"Out Interface.\" The rows are grouped into two sections. The top section is bracketed and labeled \"Network Node Segment Ids,\" listing entries such as \"L1\" mapping to \"L1\" and \"Intf1,\" \"L2\" mapping to \"L2\" and \"Intf1,\" down to \"L8\" and \"L9.\" The bottom section is bracketed and labeled \"Node Adjacency Segment Ids,\" listing entries like \"L10\" mapping to \"Pop\" and \"Intf2,\" down to \"Ln\" mapping to \"Pop\" and \"Intf5.\" To the right of this table, red text repeats the phrase \"Forwarding table remains constant,\" with a red bracket visually linking this text to the entire table structure.\n\nAt the very bottom left corner, the \"Cisco Live!\" logo appears in blue and teal. The footer at the bottom right contains the text \"BRKSP-2551,\" \"© 2023 Cisco and/or its affiliates. All rights reserved. Cisco Public,\" and the slide number \"55\" in small gray font.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/1c2e57d6-51a5-4d53-ab45-fcfb72b720c9_input_0.jpg", "images/multi_image/1c2e57d6-51a5-4d53-ab45-fcfb72b720c9_input_1.jpg", "images/multi_image/1c2e57d6-51a5-4d53-ab45-fcfb72b720c9_input_2.jpg"], "target_image": "images/multi_image/1c2e57d6-51a5-4d53-ab45-fcfb72b720c9_target.jpg", "width": 2048, "height": 1152}
+{"index": 54, "id": "59549c39-2b3b-4fde-bef6-9e6423e580ab", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean white background and is titled in large, bold blue sans-serif font at the top left: \"IPv4 for Internet Exchange Points.\" Directly below the title, a black bullet point introduces the first policy statement: \"New IXPs are delegated a /26 IPv4 address by default.\"\n\nCentered in the upper-middle portion of the slide is a two-column table designated as sub-image 1. This table is designed with a dark blue header row containing white text, followed by two data rows with alternating background colors of light periwinkle blue and white. The header row contains two distinct column labels: the left header reads \"Number of peers\" and the right header reads \"Delegation size.\" In the first data row, the \"Number of peers\" column displays the value \"60,\" while the corresponding \"Delegation size\" column reads \"Up to /25.\" In the second data row, the \"Number of peers\" column lists \"100,\" and the adjacent \"Delegation size\" column indicates \"Up to /23.\"\n\nBeneath the table area, a second black bullet point presents another guideline: \"Can request for up to /22 if you have used 60% of your existing assignment.\" Immediately following this, an asterisked footnote appears in black text, stating: \"* 163.61.160.0/20 is reserved for smaller /25 and /26 IXP assignments.\" At the very bottom of the slide, a thin horizontal blue line spans most of the width, separating the content from the footer. On the far left of this footer is the text \"APNIC\" in black, and on the far right is a small, stylized logo consisting of blue curved lines and dots resembling a network or atomic structure.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/59549c39-2b3b-4fde-bef6-9e6423e580ab_input_0.jpg"], "target_image": "images/multi_image/59549c39-2b3b-4fde-bef6-9e6423e580ab_target.jpg", "width": 2048, "height": 1152}
+{"index": 55, "id": "1fcc0858-74a2-4f0e-b93e-85973ebc02bf", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#1E3A8A\",\n      \"secondary\": \"#FFFFFF\",\n      \"text\": \"#000000\",\n      \"accent\": \"#1E3A8A\"\n    },\n    \"special_instructions\": \"Clean, high-contrast academic slide design. The header features a solid dark blue bar spanning the full width. The body text is black on a white background. The layout is asymmetrical, with text on the left and a grid of charts on the right. The charts themselves use a white background with black grid lines and multi-colored data points.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Informative\",\n    \"composition_style\": \"Asymmetric, Two-column layout (Text Left, Visuals Right)\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 190,\n      \"blocks\": [\n        {\n          \"id\": \"header-190\",\n          \"type\": \"text\",\n          \"role\": \"header\",\n          \"content\": \"Introduction | Research | Industry | Politics | Safety | Predictions\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"minor\",\n            \"position\": \"top-left\",\n            \"container\": \"banner\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"White text on a solid dark blue (#1E3A8A) background bar. Separated by vertical bars.\"\n          }\n        },\n        {\n          \"id\": \"header-right-190\",\n          \"type\": \"text\",\n          \"role\": \"header\",\n          \"content\": \"#stateofai | 189\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"minor\",\n            \"position\": \"top-right\",\n            \"container\": \"banner\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"White text on the same dark blue header bar.\"\n          }\n        },\n        {\n          \"id\": \"title-190\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Direct Preference Optimization offers an escape from “reward hacking”...or does it?\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold title text positioned below the header bar.\"\n          }\n        },\n        {\n          \"id\": \"body-190-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"First proposed as an alternative to RLHF in 2023, DPO has no explicit reward function and comes with efficiency advantages because it doesn’t sample from a policy during training or require extensive hyperparameter tuning. Despite its novelty, the method has already been used to align Llama 3.1 and Qwen2.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"left-column, top\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Paragraph text. Starts with a large, light blue/green triangular bullet point.\"\n          }\n        },\n        {\n          \"id\": \"body-190-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"However, there are signs that the “over-optimization” that’s traditionally associated with RLHF can also happen with DPO and other kinds of direct alignment algorithms (DAAs), despite the absence of a reward model. This worsens the more models are allowed to deviate from their starting point as they learn to align with human preferences.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"left-column, middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text. Standard black circular bullet.\"\n          }\n        },\n        {\n          \"id\": \"body-190-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"This could be the result of underconstrained objectives, where the algorithm unintentionally assigns high probabilities to out-of-distribution data.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"left-column, middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text. Standard black circular bullet.\"\n          }\n        },\n        {\n          \"id\": \"body-190-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"This is inherent to DAAs, but can be partially mitigated through careful parameter tuning and increased model size.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"left-column, bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text. Standard black circular bullet.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1197_616_1870_971.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1197_616_1870_971.jpg\"\n              }\n            ],\n            \"source\": \"Figure 5: Over-optimization results for sqrt(Forward KL) vs. winrates\",\n            \"caption\": \"Figure 5: Over-optimization results for $\\\\sqrt{Forward\\\\ KL}$ vs. winrates. The top row shows the final performance after 1 epoch of training, while the second row also includes 4 intermediate checkpoints. The fitted dotted curves are scaling laws from [21] applied to DAAs, with GPT4 winrates taking the place of the gold reward model score.\"\n          },\n          \"style\": {\n            \"position\": \"right-column, vertically centered\",\n            \"image_style\": \"Grid of 6 small line charts with black axes and grid lines. Data points are colored red, green, blue, and purple. Dotted trend lines are present.\",\n            \"container\": null,\n            \"special_instructions\": \"The image is a composite of 6 smaller charts arranged in a 2x3 grid. The caption is located directly below the image grid.\"\n          }\n        },\n        {\n          \"id\": \"footer-left-190\",\n          \"type\": \"text\",\n          \"role\": \"footer\",\n          \"content\": \"S  MIT  MASS\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Three logos/text elements: Red 'S', Red 'MIT', and a circular seal with 'MASS'.\"\n          }\n        },\n        {\n          \"id\": \"footer-right-190\",\n          \"type\": \"text\",\n          \"role\": \"footer\",\n          \"content\": \"stateof.ai 2024\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold text indicating the organization and year.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/1fcc0858-74a2-4f0e-b93e-85973ebc02bf_input_0.jpg"], "target_image": "images/multi_image/1fcc0858-74a2-4f0e-b93e-85973ebc02bf_target.jpg", "width": 2048, "height": 1152}
+{"index": 56, "id": "087b5ca4-561e-4382-9848-6738e68c1992", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "Set against a vibrant, solid yellow background, this presentation slide features a clear instructional layout designed to outline the requirements for earning revenue on a video platform. In the upper corners, decorative elements resembling interlocking gears in a slightly darker yellow shade frame the content, subtly hinting at the mechanics of the program. Centered at the top in large, bold, dark blue sans-serif font is the main heading: \"YouTube Channel Monetization Criteria.\"\n\nBeneath the title, the left side of the slide presents a bulleted list of specific requirements, with each bullet point starting with a bolded keyword followed by detailed conditions. The text is black and legible against the yellow backdrop. The list reads as follows, item by item: The first point states \"Subscribers: Have at least 1,000 subscribers.\" The second point specifies \"Watch time: Have at least 4,000 hours of valid public watch time in the last 12 months.\" The third point details the short-form content requirement: \"Shorts views: Have at least 10 million valid public Shorts views in the last 90 days.\" The fourth bullet reads \"AdSense account: Have an active AdSense account linked to your YouTube channel.\" The fifth item requires \"Two-step verification: Enable two-step verification on the Google account linked to your YouTube channel.\" The sixth point emphasizes compliance: \"Policies: Follow YouTube's Community Guidelines and Monetization Policies.\" The final bullet point addresses geography: \"Location: Live in a country where the YouTube Partner Program (YPP) is active.\"\n\nOccupying the lower-right quadrant of the slide is sub-image 1, which depicts a visual metaphor for monetization. This inset image shows a modern smartphone lying diagonally across a scattered pile of United States one-hundred-dollar bills. The phone screen displays a white background featuring the iconic red play button logo in the center, symbolizing the platform discussed in the text. The currency underneath is spread out casually, reinforcing the theme of financial gain associated with meeting the listed criteria. The overall composition balances the textual information on the left with the illustrative financial graphic on the right, using the gear motifs at the top to unify the design.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/087b5ca4-561e-4382-9848-6738e68c1992_input_0.jpg"], "target_image": "images/multi_image/087b5ca4-561e-4382-9848-6738e68c1992_target.jpg", "width": 2048, "height": 1152}
+{"index": 57, "id": "c29e5bc2-e36c-4374-bc38-542eb7adaa6e", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide features a clean white background with a prominent dark blue header and footer, establishing a professional aesthetic. On the left side, the slide is titled **Introduction** in a large, bold, dark blue font. Beneath the title, the text states that **University startup accelerators have emerged as a critical pillar in the global entrepreneurship ecosystem. By combining academic excellence with real-world business support, these accelerators help transform innovative ideas into scalable startups. Designed for students, researchers, and early-stage founders, university-based accelerators provide structured programs that accelerate growth, reduce risk, and drive sustainable innovation.** A second paragraph follows, declaring: **At Innovosource, we believe university startup accelerators play a vital role in shaping tomorrow’s entrepreneurs and strengthening innovation-driven economies.** To the right of this text block, sub-image 1 is positioned, displaying a top-down photograph of three individuals seated around a desk, actively reviewing documents and looking at a mobile device.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/c29e5bc2-e36c-4374-bc38-542eb7adaa6e_input_0.jpg"], "target_image": "images/multi_image/c29e5bc2-e36c-4374-bc38-542eb7adaa6e_target.jpg", "width": 2048, "height": 1152}
+{"index": 58, "id": "fa36b1fb-6724-469b-8399-ec080ad97a79", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "The presentation slide is designed with a clean white background and features a prominent header section at the top consisting of a wide dark green banner. Inside this banner, the title \"MODES OF GIFT IN HINDU LAW\" is displayed in large, white, serif capital letters. Just above the green banner, three horizontal decorative bars in dark green, lime green, and grey add a graphical accent to the top margin.\n\nBelow the header, the slide is divided into two main columns. The left column contains the primary text content, beginning with a bold black subheading labeled \"3) CORPOREAL ACCEPTANCE\". Immediately following this header is a paragraph of standard black serif text explaining the legal concept. The text reads: \"Corporeal Acceptance Is Expressed By Physical Acts Or Gestures. The Donee Shall Physically Take Possession Of The Gifted Property. This Is To Show, On His Own Part, Express Acceptance.\" Beneath this, a separate sentence continues the definition: \"This Act Signifies That The Receiver Accepts The Gift And Is Now Ready To Acquire Ownership.\"\n\nTo the right of this text block is a rectangular illustration, identified as sub-image 1, which occupies the majority of the right-hand side of the slide. This graphic features a stylized, circuit-board-like diagram in teal and blue tones against a light beige background. At the very top of this sub-image, the text \"Acceptance of Gift Arceprte under Hindu Law\" is printed in dark blue. The diagram is anchored by three large cloud shapes at the top row, containing the white text \"Mental Acceptance,\" \"Verbal Acceptance,\" and \"Corporeal Acceptance\" respectively. These clouds are interconnected by winding teal pathways that resemble electronic traces, linking to various smaller nodes and icons. Among these icons are representations of gifts, hands, and a building, symbolizing legal concepts. At the bottom of this diagram, two blue stylized hands are raised in a gesture of acceptance. To the left of the hands is a balance scale, and to the right is a classical building with pillars. Smaller cloud nodes within the diagram repeat the phrases \"VERBAL ACCEPTANCE\" and \"CORPOREAL ACCEPTANCE\" in white text.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/fa36b1fb-6724-469b-8399-ec080ad97a79_input_0.jpg"], "target_image": "images/multi_image/fa36b1fb-6724-469b-8399-ec080ad97a79_target.jpg", "width": 2048, "height": 1152}
+{"index": 59, "id": "8e1e3a95-705c-4fb6-bdd8-b0cced672d66", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a two-part layout divided by a sharp diagonal line running from the upper left corner down towards the lower left edge, creating a distinct asymmetry. The left side is dominated by a solid, deep navy blue background, while the right side presents a clean, stark white canvas.\n\nOn the deep blue left panel, the slide's hierarchy begins with a header enclosed in a lighter, bright blue rectangular banner near the top. Inside this banner, the word **Catalysts** is written in a large, bold white sans-serif font. Below this banner, aligned to the left against the navy background, is a vertical list of three terms in a smaller, white sans-serif font. The first item reads **Inhibitors**, followed by **Nourishers** directly beneath it, and finally **Toxins** at the bottom of the list.\n\nOccupying the majority of the white right panel is a large, circular diagram that resembles a stylized hub or atomic model. This diagram is constructed with a central core and an orbiting ring. The center is a soft, gradient sphere fading from light blue at the edges to white in the very middle. Emblazoned across this central sphere in bold, dark blue uppercase letters are the words **COMMON** on the top line and **GOAL** on the line below it.\n\nSurrounding this central sphere is a thick, broken ring in shades of blue and cyan. The ring is split into two primary arcs separated by white gaps at the top and bottom. The left arc of the ring is a gradient blue and contains the white text **COURAGE TO CHALLENGE** running vertically upwards along the curve. The right arc is a gradient teal-cyan and displays the white text **COURAGE TO BE CHALLENGED** running vertically upwards as well. Bridging the top and bottom of the ring are the actions associated with the central goal. At the very top of the circle, positioned between the two ring arcs, is the word **ENABLE** in medium-blue uppercase letters. At the very bottom of the circle, similarly positioned between the arcs, is the word **EMPOWER** in the same medium-blue uppercase style.\n\nAt the very bottom right of the white section, below the circular diagram, there is a small footer line in light gray text. It reads **Agile Leadership: Small Wins – Big Impact** followed by a space and the handle **@DipeshPala**. The image provided as sub-image 1 corresponds to the visual content of the circular diagram on the right side of the slide, occupying the central and right portions of the frame.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/8e1e3a95-705c-4fb6-bdd8-b0cced672d66_input_0.jpg"], "target_image": "images/multi_image/8e1e3a95-705c-4fb6-bdd8-b0cced672d66_target.jpg", "width": 2048, "height": 1152}
+{"index": 60, "id": "320aa72a-b8fb-4e1a-9ec5-08a68d4ecf4a", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The slide, branded with \"citoollkit.com\" in the top right, features the header \"Just-In-Time\" on the left. Below this, the text explains: \"The Kanban system is part of Lean production and supports Just-In-Time (JIT) production, where the right items are made in the right quantity at the right time.\" The slide then lists two points with arrow icons: first, \"The system will prevent from having excess materials lying around,\" and second, \"Instead, you will have them when you need them, Just-In-Time to meet your production needs.\" To the right of these points is sub-image 1. The slide concludes with a centered sentence at the bottom: \"Kanban is a tool that allows you to get the materials you need in time for use.\" The footer contains \"Kanban\" on the left and the number \"11\" on the right.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/320aa72a-b8fb-4e1a-9ec5-08a68d4ecf4a_input_0.jpg"], "target_image": "images/multi_image/320aa72a-b8fb-4e1a-9ec5-08a68d4ecf4a_target.jpg", "width": 2048, "height": 1152}
+{"index": 61, "id": "3fe403f1-9601-4c59-b6df-4f4db76d8123", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "Set against a clean white background, this presentation slide, marked with the slide number 34 in the top-right corner, is designed to explore the stakeholders involved in fraud detection. A decorative header runs across the top, consisting of a dark olive-green band overlaid by a thinner, lighter lime-green strip that features a modern, slightly offset geometric cutout. Dominating the upper portion of the slide in a large, dark green sans-serif font is the central question: \"Who is Responsible for Detecting Internal Fraud?\".\n\nThe slide is visually divided into two primary sections below the header. On the left side, occupying roughly the lower half of that quadrant, is sub-image 1. This rectangular image, framed with a subtle drop shadow, features a visual representation of a detective looking intently through a magnifying glass, symbolizing the scrutiny required for fraud detection.\n\nTo the right of this image, a vertical bulleted list details the potential responsible parties. Each item in the list is preceded by a small green circular bullet point and is presented in the same dark green text color as the main title. The list reads sequentially from top to bottom: \"external auditors\", \"company accountants\", \"senior management\", \"colleagues\", \"financial authorities (e.g. FSA)\", and \"police\". The layout is spacious and uncluttered, ensuring that the white space between the bulleted items enhances readability while maintaining a professional and educational tone.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/3fe403f1-9601-4c59-b6df-4f4db76d8123_input_0.jpg"], "target_image": "images/multi_image/3fe403f1-9601-4c59-b6df-4f4db76d8123_target.jpg", "width": 2048, "height": 1152}
+{"index": 62, "id": "c7864571-b5af-42ba-bdd7-fe92c67a9395", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Industrial / Tech\",\n    \"background\": {\n      \"color\": \"#F0F4F8\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#6C63FF\",\n      \"secondary\": \"#9FA8DA\",\n      \"text\": \"#333333\",\n      \"accent\": \"#6C63FF\"\n    },\n    \"special_instructions\": \"The slide features a clean, professional aesthetic with a light grey-blue background. The design relies heavily on a specific shade of purple (#6C63FF) for headings and accents. The typography is sans-serif and modern. There is a distinct separation between the text content on the left and the large illustrative graphic on the right.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Asymmetric, Two-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 31,\n      \"blocks\": [\n        {\n          \"id\": \"title-31\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Takt Time: Matching Production to Demand\",\n          \"style\": {\n            \"color\": \"#6C63FF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left, spanning roughly 60% of the slide width\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold purple text. Serves as the main header for the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-31-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Definition\\n\\nTakt time is the rate at which a product needs to be produced to meet customer demand.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below the title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The section header 'Definition' is in bold purple (#6C63FF). The body text is standard black/grey.\"\n          }\n        },\n        {\n          \"id\": \"body-31-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"What is happen?\\n\\nTakt time (TT) > Cycle time (CT)\\n\\nTalt time < Cycle time\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below the Definition section\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The section header 'What is happen?' is in bold purple (#6C63FF). The text below lists inequalities.\"\n          }\n        },\n        {\n          \"id\": \"body-31-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Benefits\\n\\nIt helps optimize production speed, minimizes waste, and ensures a balanced flow of work.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"center column, vertically aligned with the Definition section\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The section header 'Benefits' is in bold purple (#6C63FF). The text is standard black/grey.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1010_324_1872_965.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1010_324_1872_965.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a futuristic factory assembly line with a large clock\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right side, vertically centered\",\n            \"image_style\": \"Flat vector art style, isometric perspective\",\n            \"container\": null,\n            \"special_instructions\": \"The image depicts a stylized, futuristic factory interior. Key elements include a long purple conveyor belt with white circular objects, large control panels on the left and right walls, and a prominent analog clock hanging from the ceiling in the center. The color palette of the image matches the slide's primary purple theme.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/c7864571-b5af-42ba-bdd7-fe92c67a9395_input_0.jpg"], "target_image": "images/multi_image/c7864571-b5af-42ba-bdd7-fe92c67a9395_target.jpg", "width": 2048, "height": 1154}
+{"index": 63, "id": "4c0e52d0-b42e-4e76-8fe2-8c666ebae18f", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "Titled \"Materials Used in Hybrid Kilts\" at the top right, this slide lists the key components of these garments against a beige background. It explains that the material \"Combines traditional wool with lightweight cotton or polyester\" and features \"Reinforced stitching for durability in active use.\" The text further notes that they utilize \"Breathable fabrics suitable for all weather conditions,\" equipped with \"High-quality buckles, straps, and zippers,\" and that \"Eco-friendly and sustainable options available.\" To the left, sub-image 1 is positioned in the central column, showcasing folded tartan fabrics with leather straps and sewing tools.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/4c0e52d0-b42e-4e76-8fe2-8c666ebae18f_input_0.jpg"], "target_image": "images/multi_image/4c0e52d0-b42e-4e76-8fe2-8c666ebae18f_target.jpg", "width": 2048, "height": 1152}
+{"index": 64, "id": "c287d60b-23e1-4dc0-bfb7-6fd23e62cb17", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide features a two-column layout discussing marketing strategies. The left side displays an image of a crowd of stylized human figures, where one figure in the center glows bright white with arms raised, standing out from the darker crowd behind. The right side contains text on a white background with a faint blueprint pattern. At the top, the header reads **Creating Effective Ads** followed by the sub-header **Know Your Audience**. Below this, the first bullet point begins with **Identify Your Target Market:** in bold, followed by the text: \"Tailor the ad to the specific needs, behaviors, and desires of your audience. Whether you are targeting demographics (age, gender, location) or psychographics (interests, lifestyle), understanding your audience helps ensure the message resonates.\" The second bullet point starts with **Customer Pain Points:** in bold, stating: \"Address what your audience is struggling with and offer a solution. Show them how your product or service can make their lives easier or better.\" A final bullet point elaborates: \"Knowing your audience means understanding the needs, interests, and expectations of the people you're communicating with. It's an important part of effective communication and can help you connect with and persuade your audience.\" At the very bottom of the slide, a footer credit reads: \"Search Engine Marketing by Naod Ephrem\".\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/c287d60b-23e1-4dc0-bfb7-6fd23e62cb17_input_0.jpg"], "target_image": "images/multi_image/c287d60b-23e1-4dc0-bfb7-6fd23e62cb17_target.jpg", "width": 2048, "height": 1152}
+{"index": 65, "id": "693758a0-3d0c-4dc7-ae7b-5055ace69cd1", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide features a clean, minimalist layout with text on the left and a process flowchart on the right. In the top-left corner, purple text reads **Optimising for Accessibility Standards**, positioned above the main bold title **How to AB test accessibility**. A thin gray line separates this header from the explanatory text below, which states: **Always test as there are often various ways to present an idea.**\n\nOn the right side of the slide, a vertical flowchart outlines a six-step cycle. The process begins at the top with **GATHER DATA**, moves down to **CREATE HYPOTHESES**, followed by **DEFINE CLEAR GOALS**, then **USER FOCUSED DESIGNS**. The fifth step is labeled **REVIEW VWO DATA**, featuring the VWO logo between the words. The final step at the bottom is **ASSess & CHALLENGE**. A purple line connects the final step back up to the first, indicating a continuous loop. The bottom-right corner includes the logo **YOGHURTDIGITAL x VWO** and the page number **36**.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/693758a0-3d0c-4dc7-ae7b-5055ace69cd1_input_0.jpg"], "target_image": "images/multi_image/693758a0-3d0c-4dc7-ae7b-5055ace69cd1_target.jpg", "width": 2048, "height": 1152}
+{"index": 66, "id": "38a669bf-3945-4028-b703-6b549ef89939", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide features a large header reading \"Add a Primary Image,\" with the words \"Primary Image\" highlighted in bold pink. Beneath this is a green banner containing the instruction, \"Use an image that is relevant and unique.\" The lower portion displays a screenshot of a content feed labeled \"Posts\" and \"Published by Liz,\" alongside a \"See more\" link on the right. This feed showcases three illustrated thumbnails: sub-image 1 is positioned on the left above the text \"Answering the Sticky Question \"What Did Your Last...\" dated \"September 30, 2015\"; sub-image 2 is in the center above the title \"Should I Pay A Hundred Dollars for Sex?\" dated \"September 25, 2015\"; and sub-image 3 is on the right above the text \"How Jessica Broke the Rules and Saved Her Career\" dated \"September 23, 2015.\" A logo for \"Stephanie Sammons\" appears in the bottom right corner.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/38a669bf-3945-4028-b703-6b549ef89939_input_0.jpg", "images/multi_image/38a669bf-3945-4028-b703-6b549ef89939_input_1.jpg", "images/multi_image/38a669bf-3945-4028-b703-6b549ef89939_input_2.jpg"], "target_image": "images/multi_image/38a669bf-3945-4028-b703-6b549ef89939_target.jpg", "width": 2048, "height": 1536}
+{"index": 67, "id": "7084aa07-2e20-4005-8477-edc7ee3dc3fe", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, titled \"Price and quality should go hand in hand for Latin American consumers,\" features a clean, white background with a distinct vertical layout. At the very top left, the section header \"PURSUIT OF VALUE IN LATIN AMERICA\" is displayed in a bright cyan sans-serif font, immediately above the main headline which appears in a large, bold, black font. Below the header area, the slide is divided into two primary visual columns. On the left, a large rounded-rectangle photograph—identified as sub-image 1—occupies roughly the upper half of the column, depicting a view from behind of a shopper with long dark hair, wearing a yellow long-sleeved shirt and a brown backpack, pushing a blue shopping cart down a supermarket aisle. Directly beneath this image, a dense block of black text explains the consumer shift, reading verbatim: \"Latin American consumers have shifted to smarter spending habits due recent economic crises, focusing on maximising value through discounts and private label. It is not about taking just the cheapest option, but about finding the right balance between price, quality and any other added benefit that a product or service can offer. If it comes with sustainable benefits, even better. Value-driven purchasing will remain an important factor shaping consumer behaviour in the region, moving forward.\"\n\nTo the right of the image and paragraph, three large statistical data points are stacked vertically, each presented with a large cyan percentage followed by a smaller black descriptive text. The top statistic highlights that \"48% of consumers in Latin America selected low price as an attribute they look for when buying food.\" Immediately below this, the second statistic states that \"80% of consumers in Latin America are planning to increase or maintain visits to discount stores.\" The third and final statistic in this column notes that \"42% of Latin American consumers see affordability as a barrier to sustainable products.\" Beneath these statistics, in the bottom-right area, a source citation is provided in a smaller, greyish font: \"Source: Euromonitor Voice of the Consumer: Lifestyles Survey, fielded Jan-Feb 2025 (n=6,043 and n=5,604) and Sustainability Survey, fielded Jan-Feb 2024 (n=6,042).\" The footer of the slide contains the \"Passport\" logo on the left, followed by the copyright notice \"© 2025 Euromonitor International. All rights reserved.\" in small grey text, and the slide number \"32\" in the bottom right corner. A thin cyan horizontal line runs across the bottom, accented by a blue vertical line on the right edge that forms a rounded corner.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/7084aa07-2e20-4005-8477-edc7ee3dc3fe_input_0.jpg"], "target_image": "images/multi_image/7084aa07-2e20-4005-8477-edc7ee3dc3fe_target.jpg", "width": 2048, "height": 1152}
+{"index": 68, "id": "87f854ae-6487-4ac0-b67e-4f4ad872a191", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, professional layout set against a stark white background, accented by bold, abstract geometric shapes in the corners. A thick vertical band of hot pink runs down the entire left edge, while a matching pink vertical bar lines the far right. In the upper-left corner, overlapping the pink band, is a solid blue rectangle partially covered by a smaller pink rectangle. Conversely, the bottom-right corner mirrors this arrangement with a pink rectangle overlapped by a blue one. A thin, horizontal blue line stretches across the bottom of the slide, sitting just above the lower margin.\n\nThe slide's primary content is divided into text on the left and a visual element on the right. At the top left, the heading \"Vision for Saint Paul\" is displayed in a large, bold, blue sans-serif font. Directly beneath this title is a paragraph of black text that reads: \"Kaohly envisions a city that works for everyone. She believes in governance that listens, collaborates, and delivers meaningful results. By prioritizing inclusive development, safety, and economic growth, she aims to create a Saint Paul where residents thrive and neighborhoods flourish.\" This text is left-aligned and uses a standard sentence case.\n\nFollowing the introductory paragraph, the text transitions to a bolded heading that says: \"Key Principles:\". Immediately below this, a bulleted list outlines three core tenets, each presented in black text: \"Transparency and accountability in leadership,\" followed by \"Community-centered decision-making,\" and concluding with \"Holistic solutions addressing long-term city challenges.\"\n\nPositioned to the right of this text block is sub-image 1, which occupies roughly the upper-right quadrant of the slide. This black-and-white photograph depicts a woman with dark, shoulder-length wavy hair wearing glasses and a dark blazer over a patterned top. She is seated at a large wooden desk in what appears to be a formal government chamber or legislature, with rows of empty wooden desks and chairs visible in the background. She holds a pen in her right hand, resting it on a piece of paper, while her left hand rests flat on the desk. She looks directly toward the camera with a slight, confident smile. Behind her, nameplates on the desks in the background are partially visible, adding to the formal setting.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/87f854ae-6487-4ac0-b67e-4f4ad872a191_input_0.jpg"], "target_image": "images/multi_image/87f854ae-6487-4ac0-b67e-4f4ad872a191_target.jpg", "width": 2048, "height": 1152}
+{"index": 69, "id": "2cbb5364-980a-4091-a680-4261d5cd4d14", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, minimalist design set against a stark white background, divided visually into two primary columns. The left side is dominated by a large conceptual graphic, identified as sub-image 1, which occupies the majority of the left-hand quadrant. This visual element portrays a collage-style office environment with a faint grid overlay, where silhouettes of people working at desks are juxtaposed with a high-rise building. Portions of this graphic are obscured by solid blocks of vibrant blue color, visually representing a \"gap\" or obstruction within the scene. Two small black silhouettes of birds fly above the top right of this image, adding a surreal touch.\n\nThe right column of the slide is dedicated to textual content, organized hierarchically to outline the central themes. At the very top, the main slide title is displayed in a large, dark blue sans-serif font reading \"Key Issues.\" Immediately below this, a section header in a lighter blue tone introduces the first topic: \"'The Communication Gap'.\" Following this header, three bullet points are listed, each prefaced by a short horizontal line and written in the same dark blue text as the title. The first bullet point states: \"Not clearly demonstrating the value proposition for each stakeholder group.\" The second bullet point reads: \"Ignoring concerns around cost, disruption, etc.\" The third bullet point says: \"Not tailoring solutions and messaging to specific needs and priorities for specific stakeholders.\"\n\nFurther down the right column, a second section begins with the header \"Lack of Transparency of Analysis & Benefits,\" styled in the lighter blue font to match the previous section header. Underneath this title, two additional points are listed using the dark blue text and horizontal line formatting. The first point in this section states: \"Lack of transparency limits stakeholders’ ability to make informed decisions.\" The final point on the slide is phrased as a question: \"How can benefits be verified?\"\n\nAt the very bottom left corner of the slide, aligned with the margin, the website address \"www.iesve.com\" is printed in a small, dark font. The overall layout utilizes negative space effectively to separate the abstract visual representation on the left from the structured, bulleted textual analysis on the right, creating a clear and professional informational hierarchy.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/2cbb5364-980a-4091-a680-4261d5cd4d14_input_0.jpg"], "target_image": "images/multi_image/2cbb5364-980a-4091-a680-4261d5cd4d14_target.jpg", "width": 2048, "height": 1152}
+{"index": 70, "id": "c79232f8-273f-468b-9edf-13b625cc41c8", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The slide features a cream background with the prominent orange header \"3. PHYSICAL PRESENCE & ACCOUNTABILITY\" centered at the top. Below this, a yellow rectangular text box on the left side displays the paragraph: \"The employees within our offshore BPO sites work in monitored office environments, which increases accountability and reduces the anonymity that can embolden malicious actors in purely remote settings. This physical presence allows for direct supervision and adherence to company policies and security protocols within a controlled workspace, a significant contrast to the often-unregulated environments of individual remote workers.\" Complementing the text, sub-image 1 is positioned on the right side of the slide, showing four professionals gathered around a white desk focused on a laptop and documents.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/c79232f8-273f-468b-9edf-13b625cc41c8_input_0.jpg"], "target_image": "images/multi_image/c79232f8-273f-468b-9edf-13b625cc41c8_target.jpg", "width": 2048, "height": 1152}
+{"index": 71, "id": "17b55bcf-9cf7-44b4-9be4-bb2f7bc73cbb", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Professional\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#003399\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#1F2937\",\n      \"accent\": \"#003399\"\n    },\n    \"special_instructions\": \"Clean, white background with high contrast. The design relies on realistic device mockups (iPhone 5, iPad) to frame the data. A vertical blue bar on the far left acts as a subtle design element. The charts inside the devices use a simple line graph style with a dark blue line and black text.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Data-Driven\",\n    \"composition_style\": \"Asymmetric, Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 14,\n      \"blocks\": [\n        {\n          \"id\": \"title-14\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Mobile Engagement Continues to Grow – As Talent Expects More Mobile Functionality\",\n          \"style\": {\n            \"color\": \"#1F2937\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold sans-serif font. Positioned in the upper left quadrant.\"\n          }\n        },\n        {\n          \"id\": \"illustration-14-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A vertical blue rectangular bar located on the far left edge of the slide, extending from the top margin down to about 15% of the slide height.\",\n          \"style\": {\n            \"illustration_style\": \"decorative shape\",\n            \"color_scheme\": \"Solid blue (#003399)\",\n            \"position\": \"far-left edge, top-aligned\",\n            \"special_instructions\": \"Acts as a visual anchor or brand element.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_218_188_654_1133.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_218_188_654_1133.jpg\"\n              }\n            ],\n            \"source\": \"Smartphone Ownership chart inside an iPhone mockup\",\n            \"caption\": \"Image: © Dreamstime.com\"\n          },\n          \"style\": {\n            \"position\": \"left-center\",\n            \"image_style\": \"Realistic device mockup with reflection\",\n            \"container\": null,\n            \"special_instructions\": \"The chart inside shows a line graph rising from 69% to 84% between 2013 and 2015.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_926_211_1906_846.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_926_211_1906_846.jpg\"\n              }\n            ],\n            \"source\": \"Tablet Ownership chart inside an iPad mockup\",\n            \"caption\": \"Image: © Dreamstime.com\"\n          },\n          \"style\": {\n            \"position\": \"right-center\",\n            \"image_style\": \"Realistic device mockup with reflection\",\n            \"container\": null,\n            \"special_instructions\": \"The chart inside shows a line graph rising from 37% to 60% between 2013 and 2015.\"\n          }\n        },\n        {\n          \"id\": \"body-14-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"73% will leave a non-mobile optimized site and 38% view your firm more negatively if site isn’t mobile friendly\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Body text located below the tablet image. Standard weight, dark grey/black color.\"\n          }\n        },\n        {\n          \"id\": \"illustration-14-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"The inavero logo located in the bottom-left corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"Logo\",\n            \"color_scheme\": \"Black text with a stylized feather icon\",\n            \"position\": \"bottom-left corner\",\n            \"special_instructions\": \"Small logo placement.\"\n          }\n        },\n        {\n          \"id\": \"illustration-14-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"The CareerBuilder logo located in the bottom-right corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"Logo\",\n            \"color_scheme\": \"Blue and black text with a globe icon\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Small logo placement.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/17b55bcf-9cf7-44b4-9be4-bb2f7bc73cbb_input_0.jpg", "images/multi_image/17b55bcf-9cf7-44b4-9be4-bb2f7bc73cbb_input_1.jpg"], "target_image": "images/multi_image/17b55bcf-9cf7-44b4-9be4-bb2f7bc73cbb_target.jpg", "width": 2048, "height": 1152}
+{"index": 72, "id": "4e707dd2-52cb-408f-9c86-7028efef6f3d", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, numbered 14 in the bottom-right corner, features a clean white background dominated by a bold blue and yellow color scheme. In the top right corner, the logo for \"BORDERLESS ACCESS\" is displayed in blue text alongside a stylized blue wave icon. The slide is split into two main sections: a visual element on the left and a textual content block on the right.\n\nOn the left side, a large, rectangular photograph with rounded corners is set against a tilted, abstract blue shape that serves as a dynamic backdrop. This photograph, identified as sub-image 1, occupies the majority of the left quadrant and depicts a professional office scene involving two men collaborating at a desk. To the right of this visual, the headline reads \"Ready to Grow Loyalty That Lasts?\" in a large, bold, blue sans-serif font.\n\nImmediately beneath the headline, the introductory phrase \"Let Us Help You Uncover:\" appears in a smaller blue font. Following this is a bulleted list of three key value propositions, each marked by a blue right-pointing arrow icon. The first bullet point states, \"Who your Core and Fringe Loyalists are.\" The second bullet point reads, \"What drives your brand affinity.\" The third and final bullet point says, \"How to convert trust into advocacy and usage into evangelism.\" All list items are written in a standard black font.\n\nAt the bottom right of the content area, beneath the bulleted list, three rectangular call-to-action buttons are aligned horizontally. These buttons are a vibrant yellow color with black text. From left to right, the first button is labeled \"Contact Us,\" the middle button reads \"Read the Blog,\" and the rightmost button displays the text \"Read the Cas e Study,\" notably splitting the word \"Case\" across two lines with \"Cas\" on the top and \"e Study\" below it. The overall layout balances the professional imagery on the left with clear, actionable text and navigational links on the right, utilizing the blue and yellow palette to draw attention to key headers and interactive elements.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/4e707dd2-52cb-408f-9c86-7028efef6f3d_input_0.jpg"], "target_image": "images/multi_image/4e707dd2-52cb-408f-9c86-7028efef6f3d_target.jpg", "width": 2048, "height": 1152}
+{"index": 73, "id": "74c7d410-dc59-4e82-abc5-9310d57122ed", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide, part of a series titled \"7-4 Kinetic Energy and the Work-Energy Principle (8 of 8),\" features \"Example 7-10: A compressed spring.\" The main text explains that \"A horizontal spring has spring constant $ k = 360 \\, \\text{N/m} $,\" followed by part (a) asking: \"How much work is required to compress it from its uncompressed length $ (x = 0) $ to $ x = 11.0 \\, \\text{cm} $?\" Part (b) continues the problem: \"If a 1.85-kg block is placed against the spring and the spring is released, what will be the speed of the block when it separates from the spring at $ x = 0 $?\" The text specifies to \"Ignore friction.\" Finally, part (c) instructs to \"Repeat part (b) but assume that the block is moving on a table and that some kind of constant drag force (friction) $ F_{D} = 2.0 \\, \\text{N} $ is acting to slow it down, such as friction (or perhaps your finger).\" To the right of the text, sub-image 1 displays a free-body diagram of the system with a block, a spring, and arrows representing forces including $ F_N $, $ mg $, $ F_S = kx $, and $ F_D $. The slide footer contains the Pearson logo and the copyright notice: \"© 2024 Pearson Education Ltd. All Rights Reserved.\"\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/74c7d410-dc59-4e82-abc5-9310d57122ed_input_0.jpg"], "target_image": "images/multi_image/74c7d410-dc59-4e82-abc5-9310d57122ed_target.jpg", "width": 2048, "height": 1536}
+{"index": 74, "id": "3d160ae7-54dc-493a-8094-be90d18fbb63", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide, titled \"Advertising Objectives for Retailers\" at the top, outlines key business goals in a bulleted list on the left side. The first point is \"Lifting short-term sales,\" followed by \"Increasing customer traffic.\" The list continues with \"Developing and/or reinforcing a retail image,\" and then \"Informing customers about goods and services and/or company attributes.\" The fifth objective listed is \"Easing the job for sales personnel,\" and the final item is \"Developing demand for private brands.\" To the right of this text is a sub-image 1 featuring a top-down view of hands collaborating around a whiteboard covered in orange circles, a megaphone icon, and the large word \"ADVERTISING\" written in black brush strokes.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/3d160ae7-54dc-493a-8094-be90d18fbb63_input_0.jpg"], "target_image": "images/multi_image/3d160ae7-54dc-493a-8094-be90d18fbb63_target.jpg", "width": 2048, "height": 1152}
+{"index": 75, "id": "26e5a751-243b-4635-baba-7954ef041203", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide features a clean white layout with a large light gray number \"5\" at the top. Below it, the text \"SUCCESS RATES\" appears in light blue. The main content states that \"Thanks to their inbound social selling program, IBM increased their sales by 400%\", with \"400%\" emphasized in bold, followed by the citation \"(IBM)\" in gray italics. To the right sits the distinctive IBM logo enclosed in a gray circle, which also appears as sub-image 1. Four solid light blue horizontal bars are aligned on the left side, balancing the composition.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/26e5a751-243b-4635-baba-7954ef041203_input_0.jpg"], "target_image": "images/multi_image/26e5a751-243b-4635-baba-7954ef041203_target.jpg", "width": 2048, "height": 1536}
+{"index": 76, "id": "7243dab9-7f88-461e-bcad-31c7cebad225", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, titled \"SAR of Barbiturates\" in a centered header box, explores the structure-activity relationships of barbiturate drugs against a background featuring a soft gradient that transitions from a dusty rose or light mauve at the top to a pale turquoise or seafoam green at the bottom. Dominating the visual center is a chemical structure diagram, identified as sub-image 1, which occupies the middle section of the slide. This diagram illustrates the basic barbiturate ring structure with atoms and bonds labeled, specifically showing \"HN\" and \"NH\" groups at the nitrogen positions and carbonyl oxygens (\"O\") at the top and sides, with an unsubstituted methylene bridge at the bottom.\n\nSurrounding this central sub-image are four distinct text blocks, each enclosed in a faint rectangular outline and containing detailed bullet points in a serif font. Starting in the upper-left quadrant, the text box is headed by the bold phrase \"N1 & N3 Modifications:\". Below this, the first bullet point states, \"Typically, hydrogen remains at N1 and N3 for activity.\" Immediately following is the second point: \"Alkylation at N1 or N3 reduces activity.\"\n\nDirectly to the right of the central chemical structure, in the upper-right quadrant, another text box addresses \"Substitutions at C-2 & C-4 (Oxygen vs. Sulfur):\". This section notes that \"Oxygen at C-2 or C-4 → Standard barbiturates (e.g., Phenobarbital).\" The subsequent point highlights that \"Sulfur at C-2 or C-4 (Thiobarbiturates) → More lipophilic, faster onset, shorter duration (e.g., Thiopental).\" A thick black arrow connects the central sub-image to this text box, indicating the specific atoms being discussed.\n\nMoving to the lower-left quadrant, a text block titled \"Lipophilicity & CNS Activity:\" explains the pharmacokinetic implications of chemical properties. The first point reads, \"Higher lipophilicity → Rapid CNS penetration → Faster onset, shorter duration (e.g., Thiopental).\" The second point contrasts this with, \"Lower lipophilicity → Slower CNS penetration → Longer duration (e.g., Phenobarbital).\"\n\nFinally, the bottom-right quadrant contains the largest text block, labeled \"Substitutions at C-5:\". A thick black arrow points downward from the bottom of the central sub-image to this box. The text explains that \"The nature and size of alkyl or aryl groups at C-5 determine potency, duration, and onset of action.\" It further details that \"Shorter alkyl chains (e.g., methyl, ethyl) → Shorter onset, shorter duration (ultra-short acting).\" The final bullet point specifies that \"Branched or bulky groups (e.g., sec-butyl, isopentyl) → Increase lipophilicity → Faster onset & shorter duration.\" All text elements are presented in black, with key terms emphasized through bolding, ensuring high contrast against the light background.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/7243dab9-7f88-461e-bcad-31c7cebad225_input_0.jpg"], "target_image": "images/multi_image/7243dab9-7f88-461e-bcad-31c7cebad225_target.jpg", "width": 2048, "height": 1152}
+{"index": 77, "id": "9e870aab-e82b-4189-a774-fe047b7b6adf", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a minimalist design set against a solid, pale cream-colored background. The layout is horizontally divided into two primary sections: a visual portrait on the left and a textual block on the right. On the left side of the slide, positioned roughly in the vertical center, is a black-and-white ink sketch of a man from the 17th century. This figure, identified by the text below him, wears a high-collared ruff and period-appropriate doublet with a visible button placket. This illustration corresponds to sub-image 1, which occupies the left-hand portion of the slide. Directly beneath this sketch, in a clean, black sans-serif font, the name \"JOHN LOCKE\" is displayed in all capital letters.\n\nTo the right of the portrait and name, a large rectangular text box dominates the space. This box has a light gray background and rounded corners, creating a soft contrast with the cream slide background. Inside this gray container, black text is left-aligned and presented in a standard sans-serif font. The text is broken into six distinct lines, presenting a philosophical concept. Reading from top to bottom, the text within the gray box states: \"at this point he propose\", followed on the next line by \"that we can use the power\", then \"of reasoning to gain\", \"knowledge and use it to\", and finally \"understand experiences\". The typography is uniform throughout the box, with no bolding or italics applied to the text, ensuring a straightforward and readable presentation of the idea attributed to the figure on the left. There are no other decorative elements, borders, or graphics on the slide, maintaining a focused, academic aesthetic.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/9e870aab-e82b-4189-a774-fe047b7b6adf_input_0.jpg"], "target_image": "images/multi_image/9e870aab-e82b-4189-a774-fe047b7b6adf_target.jpg", "width": 2048, "height": 1152}
+{"index": 78, "id": "920b815f-6e40-497f-861b-c5091f10d5a9", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a cream-colored background adorned with faint, technical blueprint-style line drawings of geometric shapes and compass roses in the corners, giving it an academic yet slightly antiquated aesthetic. Dominating the top center is the slide's main title in a large, bold serif font: \"The Modern Worldview Makes You Powerless.\" Directly beneath the title, a two-sentence introductory text block reads: \"Worldview #1, which seems disempowering, actually inspires us to live life to the fullest. Worldview #2, which promises total control, is a trap designed to enslave you.\"\n\nThe slide's layout divides the body content into two distinct columns. On the left, under the bold heading \"The Trap Explained,\" a bulleted list outlines the negative impacts of the modern worldview. Each bullet point is accompanied by a small, monochromatic sketch. The first bullet reads: \"It Isolates You: By teaching you that the source of all your problems is within you, it makes you incapable of collective action—the only way to create real change.\" Beside this text is the first sub-image (sub-image 1), a small icon appearing as a figure inside a sphere. Below this, the second bullet point states: \"It Creates Dependence: It teaches you to rely on external authority (psychiatrists, drugs) rather than your own resilience and community (friends, nature).\" This is flanked by the second sub-image (sub-image 2), depicting hands exchanging money. The third and final bullet point reads: \"It Lowers Your Ambition: The ancient goal was Eudaimonia (human flourishing; being the best you can be). The modern goal is Pleasure (just not feeling sad). We have traded greatness for comfort.\" Beside this text sits the third sub-image (sub-image 3), illustrating a crowned figure next to a sleeping figure with \"zz\" symbols.\n\nThe right side of the slide is dominated by a large, complex, and colorful illustration (sub-image 4) that visually synthesizes the slide's themes. This artwork depicts the side profile of a human head in a vintage engraving style, with the interior of the skull exposed. Inside the head, the brain is replaced by a steampunk-like industrial machinery setup, complete with brass gears, tubes, and a glowing furnace, symbolizing the mechanistic or industrial nature of the modern mind. Contrastingly, the right side of the illustration shows a group of people holding hands in a circle, looking toward a radiant, golden cathedral-like structure that emits beams of light, suggesting community, spirituality, or a higher purpose that the machinery is failing to reach. The entire composition serves as a visual metaphor for the conflict between the isolated, mechanical \"trap\" of the modern worldview and the connected, flourishing state it supposedly replaces.\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/920b815f-6e40-497f-861b-c5091f10d5a9_input_0.jpg", "images/multi_image/920b815f-6e40-497f-861b-c5091f10d5a9_input_1.jpg", "images/multi_image/920b815f-6e40-497f-861b-c5091f10d5a9_input_2.jpg", "images/multi_image/920b815f-6e40-497f-861b-c5091f10d5a9_input_3.jpg"], "target_image": "images/multi_image/920b815f-6e40-497f-861b-c5091f10d5a9_target.jpg", "width": 2048, "height": 1143}
+{"index": 79, "id": "35a61f03-86ba-4dec-8d0d-97b9057a61e9", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This slide, titled \"Labeling & Organization for Efficiency\" in large olive-green text at the top, presents three paragraphs of advice on the left against a white background. The first paragraph begins with the statement, \"A well-organized move is a successful move. Labeling your boxes clearly can make unpacking much easier and more efficient.\" It continues, \"Use a permanent marker to write the contents of each box on the side, along with the room it belongs to. This simple step can save you a lot of time and frustration when settling into your new home.\" Below this, a second paragraph reads, \"Color-coding your labels can add another layer of organization. Assign a different color to each room, and use matching colored tape or stickers on the boxes. This visual cue can help movers quickly identify where each box should go, streamlining the unloading process.\" The third paragraph advises, \"Don't forget to mark boxes containing fragile items. Use bright colors or specific labels to indicate these boxes need special handling. This precaution will remind everyone involved to be extra careful, reducing the risk of accidental damage.\" To the right of the text, sub-image 1 displays a white cardboard box labeled \"packeze Custom Prints\" filled with hangers, with text on its side listing \"Banners, business cards, boxes, signs, stickers and much more!\" Green arrows indicate dimensions of \"20\"\" and \"24\"\".\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/35a61f03-86ba-4dec-8d0d-97b9057a61e9_input_0.jpg"], "target_image": "images/multi_image/35a61f03-86ba-4dec-8d0d-97b9057a61e9_target.jpg", "width": 2048, "height": 1152}
+{"index": 80, "id": "eba36831-a5c4-4a21-89d6-b65d23b3dfb5", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "A slide featuring a black background is visually anchored on the left by a colorful stained-glass window design, referred to as sub-image 1. To the right of this illustration, white text sets the tone with the opening line: \"I confess to almighty God and to you, my brothers and sisters, that I have greatly sinned, in my thoughts and in my words, in what I have done and in what I have failed to do, through my fault, through my fault, through my most grievous fault; therefore I ask blessed Mary ever-Virgin, all the Angels and Saints, and you, my brothers and sisters, to pray for me to the Lord our God.\"\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/eba36831-a5c4-4a21-89d6-b65d23b3dfb5_input_0.jpg"], "target_image": "images/multi_image/eba36831-a5c4-4a21-89d6-b65d23b3dfb5_target.jpg", "width": 2048, "height": 1152}
+{"index": 81, "id": "7c13e573-943a-4f8d-b536-51a56945db8d", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, professional layout set against a stark white background, with a thin, solid gold horizontal bar running across the very bottom edge. The slide is divided into two primary sections: a text-heavy left column and a visual right column containing a photograph. At the top left, a bold, gold-colored heading reads \"Across Borders.\" Immediately below this title, a paragraph of black sans-serif text explains the advantages of citizenship, stating: \"Portuguese passports for Goans allow them to have the freedom to move freely across the European Union, gaining the right to live, work, and travel. The biggest benefit of Portuguese citizenship is that you can now do all this without bureaucracy with visas, immigration problems, airport checks, and deadlines.\"\n\nFollowing this main paragraph, a shorter, separated sentence sits on its own line: \"You can also travel to other countries like the USA, Canada, and Japan without needing a visa.\" Below this section is a second major heading, styled in the same gold font as the first, which reads \"Improved Academic Life.\" The text beneath this header details educational benefits, beginning with the sentence: \"A significant benefit is access to education.\" The paragraph continues with the following verbatim text: \"You get free or subsidised access to most public institutions in Portugal or across the EU.\" The paragraph concludes with a statement where the phrase \"benefits of immigration to Portugal for\" is highlighted in blue and underlined, indicating a hyperlink: \"This is one of the most impactful benefits of immigration to Portugal for Goans.\"\n\nAt the bottom left of the text column, the website address \"www.pnfg.in\" is displayed in a standard gray font. On the right side of the slide, positioned to the right of the text columns, is a photograph (sub-image 1) that occupies roughly the middle vertical third of the slide. The image depicts a scene centered around a white table where a person is signing a document on a blue clipboard using a silver pen. In the foreground on the table, a set of house keys is visible next to the clipboard. In the background, slightly out of focus, stands another individual wearing a light green shirt. The photograph is overlaid with a soft, subtle drop shadow to give it depth, and the overall color palette of the slide is restrained, utilizing white, black, and gold to maintain a formal and clear aesthetic.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/7c13e573-943a-4f8d-b536-51a56945db8d_input_0.jpg"], "target_image": "images/multi_image/7c13e573-943a-4f8d-b536-51a56945db8d_target.jpg", "width": 2048, "height": 1152}
+{"index": 82, "id": "58fd278a-971d-434a-8ab5-73d220b0f2da", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This slide from Rubixe, which features the tagline \"Disruptive Technologies at Work\" in the top right, is titled \"Data Quality & Accuracy KPIs.\" A list of four definitions outlines key metrics: \"Data Accuracy ensures correctness of information,\" followed by \"Data Completeness checks for missing values.\" The text continues with \"Data Consistency tracks uniformity across systems,\" and concludes with \"Data Timeliness measures how up-to-date data is.\" To the right of these points, sub-image 1 is positioned in the center-right area, displaying a gold seal with a checkmark alongside a blue database icon topped with a bar chart. At the bottom of the slide, the website \"www.rubixe.com\" is displayed.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/58fd278a-971d-434a-8ab5-73d220b0f2da_input_0.jpg"], "target_image": "images/multi_image/58fd278a-971d-434a-8ab5-73d220b0f2da_target.jpg", "width": 2048, "height": 1152}
+{"index": 83, "id": "dbd95fb5-fe26-4e8a-85db-c628c22587ec", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, minimalist layout on a white background, designed to instruct the viewer on incorporating engagement tactics into their writing. Dominating the top center is the title \"Insert a Call-to-Action,\" where \"Insert a\" is rendered in a bold, dark grey sans-serif font, while \"Call-to-Action\" immediately follows in a large, vibrant pink, playful font. Below the title, the slide is divided into two primary visual zones. On the left side, occupying roughly the lower-left third of the slide, is a solid, muted blue-grey rectangular block. Inside this block, white sans-serif text explains the core concept: \"At the end of your post, include a call-to-action! (can include a hyperlinked image).\"\n\nTo the right of the blue block and slightly lower, a grey-bordered rectangle contains a practical example. This box displays a 3D render of a book cover as a sub-image, positioned on the left side of the frame. The sub-image shows a dark blue book titled \"TransForm\" with the subtitle \"Dramatically Improve Your Career, Business, Relationships, and Life ...\" and a red banner reading \"ONE SIMPLE STEP AT A TIME\". The author's name, \"Jeff Haden,\" appears at the bottom, and the phrase \"Click the image\" is superimposed in red text over the cover. Beside this book image, three paragraphs of text provide a specific copywriting example. The first paragraph reads, \"Check out my book of personal and professional advice, TransForm: Dramatically Improve Your Career, Business, Relationships, and Life - - One Simple Step At a Time. (PDF version here, Kindle version here, Nook version here.)\" The specific links and the title are styled in a blue color to indicate hyperlinks. The second paragraph follows with a guarantee: \"If after 10 minutes you don't find at least 5 things you can do to make your life better I'll refund your money.\" The final line concludes with the sentiment, \"That way you have nothing to lose... and everything to gain.\" In the bottom right corner of the slide, outside the grey box, is a logo consisting of a pink circle containing the white initials \"SS\" next to the name \"Stephanie Sammons\" in green text.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/dbd95fb5-fe26-4e8a-85db-c628c22587ec_input_0.jpg"], "target_image": "images/multi_image/dbd95fb5-fe26-4e8a-85db-c628c22587ec_target.jpg", "width": 2048, "height": 1536}
+{"index": 84, "id": "69a27229-79d0-4bd2-9a1f-3bc732f81950", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "Set against a solid, dark charcoal gray background that covers the entire frame, the slide presents a clear call to action centered at the top in white sans-serif text reading, \"Share this eBook with your friends:\". Directly below this headline, four circular icons are arranged horizontally in equal spacing, each encased in a white border. The first icon on the left features a light blue field containing the lowercase letter 'f', with the text \"FACEBOOK\" printed in orange uppercase letters centered beneath it. Moving right, the second icon mirrors the design with a blue circle containing a white silhouette of a bird in flight, labeled underneath with \"TWITTER\" in the same orange uppercase font. The third icon displays the lowercase letters \"in\" in white against a blue background, with the caption \"LINKEDIN\" appearing below in orange. The fourth and final icon on the right shows a white downward-pointing arrow within a blue circle, indicating a download function, with the word \"DOWNLOAD\" printed in orange underneath. In the bottom left corner, the text \"We Are Social\" appears in small white font, while the bottom right corner displays the footer text \"Social Brands: The Future Of Marketing • @wearesocialsg • 123\" in small white text. Each of the four circular icons corresponds to a specific sub-image provided in the source data: the Facebook icon is represented by sub-image 1, the Twitter icon by sub-image 2, the LinkedIn icon by sub-image 3, and the download arrow icon by sub-image 4, each occupying a distinct position within the central horizontal alignment of the slide.\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/69a27229-79d0-4bd2-9a1f-3bc732f81950_input_0.jpg", "images/multi_image/69a27229-79d0-4bd2-9a1f-3bc732f81950_input_1.jpg", "images/multi_image/69a27229-79d0-4bd2-9a1f-3bc732f81950_input_2.jpg", "images/multi_image/69a27229-79d0-4bd2-9a1f-3bc732f81950_input_3.jpg"], "target_image": "images/multi_image/69a27229-79d0-4bd2-9a1f-3bc732f81950_target.jpg", "width": 2048, "height": 1536}
+{"index": 85, "id": "8bb5ff03-b378-4e06-af4c-2c86cdf32ecd", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Social Media Marketing\",\n    \"background\": {\n      \"color\": \"#E0E0E0\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#545454\",\n      \"secondary\": \"#545454\",\n      \"text\": \"#545454\",\n      \"accent\": \"#545454\"\n    },\n    \"special_instructions\": \"The slide features a clean, minimalist design with a light grey background. It utilizes large, semi-transparent dark grey circles as decorative background elements in the corners. The layout is asymmetrical, balancing text on the left with two distinct circular image cutouts on the right and top-left.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Informative, Modern\",\n    \"composition_style\": \"Asymmetric, Split Layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 5,\n      \"blocks\": [\n        {\n          \"id\": \"title-5\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"CHOOSING THE RIGHT PLATFORM\",\n          \"style\": {\n            \"color\": \"#545454\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, uppercase sans-serif font. Centered horizontally.\"\n          }\n        },\n        {\n          \"id\": \"body-5-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Facebook: Best for ads and community building.\\n• Instagram: Great for visual content and younger audiences.\\n• LinkedIn: Best for professional networking and B2B marketing.\\n• Twitter (X): Ideal for real-time updates and trending topics.\\n• TikTok: Perfect for short-form video content.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#545454\",\n            \"importance\": \"supporting\",\n            \"position\": \"center-left, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point list. Standard sans-serif font. Left-aligned relative to the text block.\"\n          }\n        },\n        {\n          \"id\": \"illustration-5-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large, semi-transparent dark grey circle located in the bottom-left corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"decorative background shape\",\n            \"color_scheme\": \"dark grey\",\n            \"position\": \"bottom-left corner\",\n            \"special_instructions\": \"Semi-transparent overlay used for visual balance.\"\n          }\n        },\n        {\n          \"id\": \"illustration-5-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large, semi-transparent dark grey circle located in the top-right corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"decorative background shape\",\n            \"color_scheme\": \"dark grey\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"Semi-transparent overlay used for visual balance.\"\n          }\n        },\n        {\n          \"id\": \"illustration-5-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large, semi-transparent dark grey circle located in the bottom-right corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"decorative background shape\",\n            \"color_scheme\": \"dark grey\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Semi-transparent overlay used for visual balance.\"\n          }\n        },\n        {\n          \"id\": \"image-5-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_0_0_433_418.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_0_0_433_418.jpg\"\n              }\n            ],\n            \"source\": \"Close-up of smartphone app icons (Facebook, Instagram, WhatsApp, Google)\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-left corner\",\n            \"image_style\": \"cropped into a circle\",\n            \"container\": null,\n            \"special_instructions\": \"The image is cropped into a perfect circle and placed in the top-left area.\"\n          }\n        },\n        {\n          \"id\": \"image-5-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1561_672_2047_1150.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1561_672_2047_1150.jpg\"\n              }\n            ],\n            \"source\": \"Woman using a smartphone surrounded by social media reaction icons\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-right corner\",\n            \"image_style\": \"cropped into a circle\",\n            \"container\": null,\n            \"special_instructions\": \"The image is cropped into a perfect circle and placed in the bottom-right area.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/8bb5ff03-b378-4e06-af4c-2c86cdf32ecd_input_0.jpg", "images/multi_image/8bb5ff03-b378-4e06-af4c-2c86cdf32ecd_input_1.jpg"], "target_image": "images/multi_image/8bb5ff03-b378-4e06-af4c-2c86cdf32ecd_target.jpg", "width": 2048, "height": 1152}
+{"index": 86, "id": "b475ab36-14d3-43d4-a588-c6ff3638d2db", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, set against a plain white background, is titled **Reversal Theory** in large, dark blue, underlined text positioned in the upper left corner. Below the title, a bulleted list details the concepts of the theory in black sans-serif font. The first bullet point states: \"Reversal theory aims to discuss how performers arousal and psychological state can fluctuate from moment to moment, depending on them and the meaning they attach to a given situation at a given time. For example, a performer sometimes may be inspired by the crowd to perform better and in other situations, the pressure from the crowd may make them feel anxious.\" The second bullet point reads: \"It suggests that performers can fluctuate from relaxed to anxious and from bored to excited at various times throughout their performance. For example, a footballer may be excited at the start of a game but become more bored if they are not seeing any of the ball or if their team are winning by a lot of goals.\" The final bullet point concludes with: \"It suggests that performance is best when performers are not at either of the two extreme ends of the two spectrums and that performance is therefore at its best when athletes are in an “optimum work zone” where they are not too relaxed or anxious and not too excited or bored.\"\n\nTo the right of this text is a diagram, referenced as **sub-image 1**, occupying roughly the right third of the slide. This sub-image displays a graph with a horizontal axis labeled \"Arousal Level\" ranging from \"Low\" on the left to \"High\" on the right. The vertical axis is labeled \"Hedonic Tone,\" with \"Pleasant\" at the top and \"Unpleasant\" at the bottom. Two curved black lines form an X shape across the graph. The line starting at the top left is labeled \"Relaxed\" and curves down to the bottom right, labeled \"Anxious.\" The second line starts at the bottom left, labeled \"Bored,\" and curves up to the top right, labeled \"Excited.\" In the center of the graph, where the two lines intersect, there is a shaded yellow-orange square labeled \"Optimum Work Zone.\"\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/b475ab36-14d3-43d4-a588-c6ff3638d2db_input_0.jpg"], "target_image": "images/multi_image/b475ab36-14d3-43d4-a588-c6ff3638d2db_target.jpg", "width": 2048, "height": 1152}
+{"index": 87, "id": "2ca99ef6-a90a-49d4-b2c7-10a3792b7b45", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide, titled \"MOBILE USE\" in large green letters at the top against a dark gray background, displays statistics dated \"JUL 2014\" in a green box on the left and features a national flag icon on the right. The slide is divided into four columns. The first column is labeled \"TOTAL NUMBER OF ACTIVE MOBILE SUBSCRIPTIONS\" in white text, positioned above sub-image 1 in the left-center area; below this, the value \"39.5 M\" appears in large orange font. The second column displays the header \"PERCENTAGE OF TOTAL MOBILE SUBSCRIPTIONS THAT ARE PRE-PAID\" above sub-image 2; below this sits the figure \"91%\" in orange, with the text \"we are social\" visible near the graphic. The third column is titled \"PERCENTAGE OF TOTAL MOBILE SUBSCRIPTIONS THAT ARE POST-PAID\" and features sub-image 3 below the text, followed by the orange percentage \"9%\". The final column on the right is labeled \"PERCENTAGE OF MOBILE SUBSCRIPTIONS THAT ARE 3G CONNECTIONS\" with sub-image 4 situated below it, showing the notation \"[ N/A ]\" in gray brackets at the bottom. The footer includes the text \"We Are Social\" and \"Sources: US Census Bureau, United Nations, GSMA Intelligence\" on the left, and \"@wearesocialsg • 19\" on the right.\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/2ca99ef6-a90a-49d4-b2c7-10a3792b7b45_input_0.jpg", "images/multi_image/2ca99ef6-a90a-49d4-b2c7-10a3792b7b45_input_1.jpg", "images/multi_image/2ca99ef6-a90a-49d4-b2c7-10a3792b7b45_input_2.jpg", "images/multi_image/2ca99ef6-a90a-49d4-b2c7-10a3792b7b45_input_3.jpg"], "target_image": "images/multi_image/2ca99ef6-a90a-49d4-b2c7-10a3792b7b45_target.jpg", "width": 2048, "height": 1536}
+{"index": 88, "id": "40965cf7-fe52-459e-85fd-31b2ff803d8d", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, titled \"1. Operational Cost — API\" in a large, bold, dark sans-serif font at the top left, sets the stage for a discussion on the financial dynamics of Artificial Intelligence infrastructure. The background is stark white, providing high contrast for the text and the two large illustrative cartoons that dominate the layout. The slide is divided into a left section featuring a Sunday comic strip and a right section containing a four-panel webcomic, with a bridge of text situated in the center-left.\n\nOn the left side, occupying roughly the left third of the slide, is a full-color sub-image 1 depicting a scene from the \"Calvin and Hobbes\" comic strip. In this panel, the tiger character, Hobbes, stands behind a lemonade stand with a sign that reads \"SELF-HOSTED GPUs $1000 K\". Standing with arms crossed and a skeptical expression is the boy, Calvin. To his right, another lemonade stand is being run by a different character, with a sign that reads \"APIs $1.50\", surrounded by a crowd of smiling children.\n\nPositioned immediately to the right of this first cartoon, acting as a conceptual bridge between the two images, is a block of text describing the benefits of the option on the right. A light blue, thick arrow curves downwards and to the right above the text, visually directing the viewer's flow. The text block begins with the heading \"LLM API\" in bold. Below this, under the sub-heading \"Advantages\", is a vertical list of three items, each preceded by a green checkmark icon. The items are listed verbatim as: \"Ease of use\", \"Performance\", and \"Managed\".\n\nOn the right side of the slide, occupying the majority of the remaining space, is a black-and-white line drawing identified as sub-image 2. This four-panel comic strip is titled \"THE LLM API PRICING REALITY CHECK\" across the top, with the phrase \"API PRICING REALITY CHECK\" highlighted in a yellow marker style. The first panel shows a character labeled \"CEO\" leaning back in a chair holding a drink, telling an \"INVESTOR\" that \"So our AI costs are totally sustainable! We're only spending $1000/month on GPT calls\". The investor asks, \"Great! what's your growth plan?\". The second panel depicts the CEO surrounded by floating gears, exclaiming, \"We'll integrate AI into EVERYTHING!\". The third panel shows the investor holding up a phone and asking, \"Maybe we should've read the fine print about 'introductory pricing'\", while the investor character holds up a small card with the word \"NO\" on it. The final panel shows a shocked CEO looking at a laptop screen displaying a chart, with a speech bubble stating, \"Monthly API Bill: $47,328.50\". Beneath this comic strip is the citation \"ScaleDown - Substack\".\n\nThe bottom of the slide features branding and logos. On the bottom left, the logo \"EKRAF\" is displayed next to a blue geometric snowflake-like icon. On the bottom right, the logos for \"dicoding\" and \"bbl\" appear in dark blue text. A solid, light blue horizontal bar runs across the very bottom edge of the slide, grounding the design.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/40965cf7-fe52-459e-85fd-31b2ff803d8d_input_0.jpg", "images/multi_image/40965cf7-fe52-459e-85fd-31b2ff803d8d_input_1.jpg"], "target_image": "images/multi_image/40965cf7-fe52-459e-85fd-31b2ff803d8d_target.jpg", "width": 2048, "height": 1152}
+{"index": 89, "id": "06dd8402-9db2-4f61-9dc7-34821ded529c", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "The presentation slide features a black background with white text and a photograph of a harbor. At the top, the title reads, \"From Ponta Garça to Capelas (via Vila Franca do Campo).\" To the left is sub-image 1, which depicts a marina with boats docked under a blue sky, identified by the text as the harbor area. To the right of the image, the body text begins with the sentence, \"On August 10, my main goal was to get from Ponta Garça to Capelas, right on the other side of the island of São Miguel.\"\n\nFollowing this introduction, the slide details the route under the heading, \"The journey was as follows:\". A bulleted list follows, starting with \"Walk from Ponta Garça to Vila Franca do Campo (7.67 km, including detours),\" then \"By bus from Vila Franca do Campo to Ponta Delgada,\" and finally \"Walk from Ponta Delgada to Capelas (14.52 km).\" Below the list, the text concludes with, \"While waiting for the bus in Vila Franca do Campo, I explored the town's harbor area (in the photo).\" The slide is numbered \"46\" in the bottom right corner.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/06dd8402-9db2-4f61-9dc7-34821ded529c_input_0.jpg"], "target_image": "images/multi_image/06dd8402-9db2-4f61-9dc7-34821ded529c_target.jpg", "width": 2048, "height": 1152}
+{"index": 90, "id": "f57d8328-cf3b-4e8a-a107-1232bc20bae0", "task": "multi_image", "source": "real", "prompt_type": "short", "instruction": "This presentation slide features a soft green background split into two sections. On the left, the main title is displayed in dark green lettering, reading \"How to Travel Sustainably and Reduce Your Carbon Footprint.\" Directly below this title, the slide credits the author with the text \"By Chris Janese.\" On the right side, a rectangular illustration with rounded corners, identified as sub-image 1, depicts a hiker with a backpack and staff standing on a rocky peak while looking out over a misty mountain range and clouds.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/f57d8328-cf3b-4e8a-a107-1232bc20bae0_input_0.jpg"], "target_image": "images/multi_image/f57d8328-cf3b-4e8a-a107-1232bc20bae0_target.jpg", "width": 2048, "height": 1152}
+{"index": 91, "id": "72ece28d-268a-4906-9a4a-d28e74852dce", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide is designed with a clean, professional layout, featuring a soft, off-white background that transitions subtly into a very light gray on the right side. The slide is titled at the top left with the large, bold, black text \"Definition and Purpose.\" Dominating the right third of the slide is a large, vertically oriented photograph contained within a rounded rectangular frame. This is sub-image 1, which occupies the upper-right to lower-right portion of the slide, displaying an assortment of heavy-duty metal mechanical parts and a laptop on a dark wooden surface.\n\nBeneath the main title on the left side, there is a paragraph of body text in a smaller, dark gray sans-serif font. The text begins with the sentence: \"A heavy duty adjustable hitch allows users to customize the height and angle of their trailer connections.\" Immediately following this, the text continues in the same paragraph with: \"It is engineered to provide flexibility, enabling compatibility with various trailer sizes and types.\" The paragraph concludes with the final sentence: \"This adaptability optimizes towing performance by achieving the proper balance and alignment.\"\n\nThe visual element, sub-image 1, is positioned to the right of the text block and extends almost to the right edge of the slide. It shows a close-up view of a dark table scattered with various metallic towing hardware, including shackles, brackets, and adjustable hitches, some appearing slightly rusted or industrial. In the background of this photograph, an open laptop computer is visible, with its screen displaying rows of data or a list, suggesting a planning or engineering context. The lighting in the photograph is somewhat dramatic, highlighting the metallic textures against the dark wood. The overall color palette of the slide is neutral, relying on grays, whites, and the dark tones of the photograph to create a focused, informative aesthetic. The layout clearly separates the textual definition from the visual representation, guiding the viewer's eye from the title down the text and then to the detailed image on the right.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/72ece28d-268a-4906-9a4a-d28e74852dce_input_0.jpg"], "target_image": "images/multi_image/72ece28d-268a-4906-9a4a-d28e74852dce_target.jpg", "width": 2048, "height": 1152}
+{"index": 92, "id": "269a264d-bfc9-4570-9c17-7a57d51bed4e", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean white background topped by a solid, deep blue horizontal banner. Within this banner, the title is displayed in a white, sans-serif font: \"Model-averaging: synthetic control with 3 donor units, 2 pre-treatment periods\". Below the header, the main content area is divided into two side-by-side graphical panels. The left panel displays a coordinate system labeled with \"yi2\" on the vertical axis and \"yi1\" on the horizontal axis. Inside this graph, a large triangle connects three points labeled \"IL\" in the upper left, \"MA\" in the upper right, and \"NJ\" in the lower right. Within this triangle, a specific point labeled \"CA\" in blue is shown with dashed lines projecting to the triangle's edges. Various labels surround these points, denoting estimates like \"ŷCA{IL}\", \"ŷCA{MA,IL}\", \"ŷCA{MA}\", \"ŷCA{NJ,MA}\", \"ŷCA{NJ}\", and \"ŷCA{IL,NJ}\". A light tan shading covers the area of the large triangle. Beneath this graph is the label \"Non-interpolating\" in black text. To the right, the second panel mirrors the coordinate system and triangle structure with vertices labeled \"IL\", \"MA\", and \"NJ\". It similarly displays a central point labeled \"CA\" in blue. This graph includes labels for \"ŷCA{MA,IL}\", \"ŷCA{NJ,MA}\", \"ŷCA{IL,NJ}\", and \"ŷCA{MA,IL,NJ}\". Distinct from the left panel, a smaller triangle formed by the central \"CA\" point and two other points on the boundary is filled with a light orange shade, while the rest of the main triangle remains unshaded or white. Below this graph is the label \"Interpolating\". At the bottom of the slide, spanning the width of the content area, is a concluding statement in black text: \"Model-averaging → distribution of prediction risk → improved average performance\". In the bottom right corner, the page number \"6\" is visible.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/269a264d-bfc9-4570-9c17-7a57d51bed4e_input_0.jpg", "images/multi_image/269a264d-bfc9-4570-9c17-7a57d51bed4e_input_1.jpg"], "target_image": "images/multi_image/269a264d-bfc9-4570-9c17-7a57d51bed4e_target.png", "width": 2880, "height": 1620}
+{"index": 93, "id": "7e3803f9-48ec-4c30-a42b-e4924ef9c06f", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "Set against a clean white background, this presentation slide is titled **In the context of past approaches** in large, dark grey sans-serif font at the top. The slide is numbered **21** in the top-left corner and **96** in the bottom-right. The main body consists of three distinct 2D scatter plots arranged horizontally, each depicting a different methodology for decision-making based on predicted utilities.\n\nThe first plot on the left is labeled **True Optimization** in bold black text at the bottom. The chart's horizontal axis represents **Predicted Utility for A** and the vertical axis represents **Predicted Utility for B**, both ranging from -2 to 2 with grid lines every 0.5 units. The plot area is divided into two triangular regions by a dashed blue diagonal line rising from the bottom-left to the top-right. The upper-left triangle is shaded light blue and contains the text **DQ = 1** in bold black. The lower-right triangle is shaded light red and labeled **DQ = 0** in bold black. A yellow five-pointed star with a black outline marks a specific point in the blue region at approximately coordinates (0.5, 1.5).\n\nThe middle plot is titled **Sampled Points** in bold black text below the graph. It uses the same axes as the first plot: **Predicted Utility for A** horizontally and **Predicted Utility for B** vertically, ranging from -2 to 2. Unlike the first plot, there is no background shading or diagonal divider line. Instead, the graph displays a scatter of fifteen circular data points overlaid on the white grid. Most points are light blue circles, while three points located near the bottom-right quadrant are light red circles. The yellow star with a black outline appears in the same position as in the first plot, at approximately (0.5, 1.5). The blue dots are scattered primarily in the upper-left and central areas, while the red dots are clustered in the lower-right area.\n\nThe third plot on the right is labeled **Learned Loss** in bold black text, with a parenthetical subtitle **(Without Handcrafting)** directly beneath it. This graph replicates the layout of the first plot, featuring the same axes labels (**Predicted Utility for A** and **Predicted Utility for B**) and the same range of -2 to 2. It overlays the same fifteen data points (blue and red circles) seen in the middle plot, along with the yellow star at (0.5, 1.5). However, this plot also includes a background color gradient that mimics the \"True Optimization\" shading. The background is light blue in the upper-left and transitions smoothly to light red in the lower-right, visually suggesting a learned decision boundary that approximates the diagonal split shown in the first graph.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/7e3803f9-48ec-4c30-a42b-e4924ef9c06f_input_0.jpg", "images/multi_image/7e3803f9-48ec-4c30-a42b-e4924ef9c06f_input_1.jpg", "images/multi_image/7e3803f9-48ec-4c30-a42b-e4924ef9c06f_input_2.jpg"], "target_image": "images/multi_image/7e3803f9-48ec-4c30-a42b-e4924ef9c06f_target.png", "width": 1440, "height": 810}
+{"index": 94, "id": "e7235cf6-121e-4fc9-b0a1-4a6510e93658", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, minimalist white background that sets a professional and focused tone. At the top left corner, the title \"Future Works\" is displayed in a bold, black sans-serif font, clearly indicating the topic of the slide. Below the title, on the left-hand side, two bullet points are listed in black text, outlining specific research objectives. The first bullet point reads: \"Design permutation invariant motor neurons and achieve permutation invariance in the output space.\" Directly beneath it, the second bullet point states: \"Include environmental reward to train permutation invariant meta-learning agents that can adapt to not only changes in the environment, but also changes to itself.\" Both sentences are presented in a standard, readable font size and are left-aligned.\n\nOccupying the right side of the slide is sub-image 1, which depicts two distinct robotic figures positioned against the white void. On the left within this sub-image is a quadrupedal robot resembling a dog, colored in matte black with lighter gray legs, standing in a ready posture. To its right stands a humanoid robot with a white and light gray body and a black head unit; its right arm is raised in a waving gesture. These visual elements occupy roughly the middle-to-lower vertical band on the right side of the slide. In the bottom right corner, the \"Google Research\" logo is visible, featuring the multi-colored \"Google\" wordmark followed by \"Research\" in gray text, serving as a footer for the presentation. No other decorative lines, borders, or background textures are present, ensuring the viewer's attention remains fixed on the text and the robotic imagery.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/e7235cf6-121e-4fc9-b0a1-4a6510e93658_input_0.jpg"], "target_image": "images/multi_image/e7235cf6-121e-4fc9-b0a1-4a6510e93658_target.png", "width": 2880, "height": 1620}
+{"index": 95, "id": "346cd3c9-1185-466c-aa22-e351efdfe2b0", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, set against a clean white background, is titled \"Sparse Winning Tickets are robust to distributional shifts\" in a large, black sans-serif font at the top left. Below the title, two lines of text establish the research questions: \"Does sparsity reduce overfitting by avoiding memorization of training samples?\" followed by \"Evaluate robustness of networks on unseen distributional shifts.\" The main body of the slide is divided into two columns illustrating different types of data variations. On the left side, under the heading \"source domain\" and \"target domains,\" sits sub-image 1, which occupies the middle-left portion of the slide. Below this image group is the text \"Domain Shifted *\". A footnote at the very bottom left clarifies the asterisk: \"* Representative figure from http://ai.bu.edu/visda-2021/\". The right column features a larger visual element, sub-image 2, located in the upper-right quadrant. This sub-image displays a grid of 15 photographs of a yellow bird on a branch, each subjected to different visual distortions. The labels for these distortions are printed directly above or between the rows within the sub-image itself. The first row is labeled with \"Gaussian Noise\", \"Shot Noise\", \"Impulse Noise\", \"Defocus Blur\", and \"Frosted Glass Blur\". The subsequent rows within the image contain labels such as \"Motion Blur\", \"Zoom Blur\", \"Snow\", \"Frost\", \"Fog\", \"Brightness\", \"Contrast\", \"Elastic\", \"Pixelate\", and \"JPEG\". Beneath this large grid of corrupted images, the text reads \"Synthetic Corruptions\", followed by the citation \"(Hendrycks et al. '19)\" in a smaller font size.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/346cd3c9-1185-466c-aa22-e351efdfe2b0_input_0.jpg", "images/multi_image/346cd3c9-1185-466c-aa22-e351efdfe2b0_input_1.jpg"], "target_image": "images/multi_image/346cd3c9-1185-466c-aa22-e351efdfe2b0_target.png", "width": 1440, "height": 810}
+{"index": 96, "id": "1c057d91-5540-46d5-abcf-02dfda8b9b24", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, white background that serves as a neutral canvas for demonstrating a technical process titled \"HAT generating counterfactual videos.\" This title is positioned at the top left of the slide in a large, bold, black sans-serif font, immediately establishing the subject matter. Below this heading, the visual section is divided into two distinct areas arranged horizontally. On the left side, sub-image 1 occupies approximately 40% of the slide's width. This rectangular image depicts a scene within an indoor setting, likely a gym, featuring a blue floor and a textured rock-climbing wall in the background. Superimposed over the central figure in the scene is a solid white silhouette, effectively masking the person's identity and form while highlighting their posture in a forward-bending or handstand position. To the right of this first image, and separated by a significant amount of white space, is sub-image 2, which is noticeably smaller, taking up roughly 15% of the slide width. This image displays a figure against a completely white background, isolated from any environment, showing a person in dark clothing performing an acrobatic handstand or cartwheel motion. Below these visual examples, the slide presents two citation references in a smaller, standard black font. The first line reads: \"Jitesh Jain, et al. Semask: Semantically masking transformer backbones for effective semantic segmentation, arXiv\". Immediately following this is the second reference: \"Dahun Kim, et al. Deep video inpainting, CVPR 2019\". These citations provide the academic context for the technologies shown. At the very bottom of the slide, running the full width, is a distinct footer banner in a solid light orange color. Inside this banner, white text provides a third reference: \"Jihoon Chung, et al. “Enabling Detailed Action Recognition Evaluation Through Video Dataset Augmentation” NeurIPS 2022 Datasets and Benchmarks\". The overall layout is structured and academic, moving from a conceptual title to visual evidence of the method's application, and finally grounding the presentation in relevant scientific literature.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/1c057d91-5540-46d5-abcf-02dfda8b9b24_input_0.jpg", "images/multi_image/1c057d91-5540-46d5-abcf-02dfda8b9b24_input_1.jpg"], "target_image": "images/multi_image/1c057d91-5540-46d5-abcf-02dfda8b9b24_target.png", "width": 2880, "height": 1620}
+{"index": 97, "id": "fc26f4fb-a397-4479-a473-56e1b741a450", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "Set against a clean white background, the presentation slide is titled \"Retrieval Results\" in a large, serif font positioned at the top left. The main body of the slide is organized into two distinct columns, labeled (a) and (b), each comparing different query methods using overlapping histograms and grid-based ECG visualizations. The left column, identified by the caption \"(a) DTC query {SR, male, under 49}\" located at the bottom, begins at the top with sub-image 1. This histogram features a vertical axis labeled \"Density\" ranging from 0.00 to 1.25 and a horizontal axis labeled \"Euclidean distance\" ranging from 1.4 to 2.2. The chart displays overlapping density distributions for four categories listed in a legend titled \"class\": SR (teal), SB (orange), GSVT (blue), and AFIB (pink). Visually, the distribution is dominated by overlapping teal and orange peaks between distances 1.5 and 1.9, with blue and pink elements less distinct. Beneath this chart sits sub-image 2, which occupies the lower-left portion of the left column; it displays six rectangular panels arranged in two rows and three columns, showing ECG waveforms, all of which are framed in red borders.\n\nThe right column, captioned at the bottom as \"(b) CP CROCS query {SR, male, under 49}\", mirrors the layout of the left but presents different data distributions. At the top of this column is sub-image 3, a histogram with a vertical \"Density\" axis scaling from 0.0 to 0.4 and a horizontal \"Euclidean distance\" axis spanning from 0.5 to 3.5. The \"class\" legend remains consistent, listing SR, SB, GSVT, and AFIB in their respective colors. In this distribution, the teal SR histogram forms a distinct, sharp peak on the far left between 0.5 and 1.0, while the orange SB distribution forms a broader central peak around 2.2, and the blue GSVT appears as a separate cluster on the far right near 3.0. Below this chart is sub-image 4, located in the lower-right quadrant, presenting a 2x3 grid of ECG panels. In contrast to the left column, these panels feature a mix of colored borders: the top-left, top-right, and bottom-right panels are bordered in red, while the top-middle, bottom-left, and bottom-middle panels are bordered in green. A horizontal navigation banner spans the entire bottom of the slide in a dark blue hue. It contains six rectangular tabs labeled sequentially from left to right: \"Cardiac Data,\" \"Background,\" \"Research Question,\" \"Methods,\" \"Results,\" and \"Conclusion.\" The \"Results\" tab is highlighted in yellow, indicating the current section. On the far left of this footer is the logo for \"X lab Computational Health Informatics,\" and on the far right is a circular seal representing Imperial College London.\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/fc26f4fb-a397-4479-a473-56e1b741a450_input_0.jpg", "images/multi_image/fc26f4fb-a397-4479-a473-56e1b741a450_input_1.jpg", "images/multi_image/fc26f4fb-a397-4479-a473-56e1b741a450_input_2.jpg", "images/multi_image/fc26f4fb-a397-4479-a473-56e1b741a450_input_3.jpg"], "target_image": "images/multi_image/fc26f4fb-a397-4479-a473-56e1b741a450_target.png", "width": 1920, "height": 1080}
+{"index": 98, "id": "0cabf7c1-b7b7-4ce0-82c9-9d3c43ed9f88", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean white background dominated by technical diagrams and text concerning autonomous vehicle systems. In the top right corner, a circular logo composed of purple dots is labeled \"NEURAL INFORMATION PROCESSING SYSTEMS\". The main title, displayed prominently in the top left in large, bold, blue font, reads \"VIC3D Object Detection\". Directly below this, a subtitle in bold black text introduces \"The challenges of VIC3D Object Detection\", followed by a bulleted list identifying two specific issues: \"Limited communication bandwidth\" and \"Uncertain latency\".\n\nThe lower two-thirds of the slide are occupied by a large, composite diagram, identified as sub-image 1, which spans the full width of the content area. This diagram is divided into three vertical sections labeled (a), (b), and (c/d). On the far left, section (a) contains two vertically stacked images depicting sensor data views of a road scene, with the caption \"(a) Infrastructure and Vehicle Sensor Data\" located below. The center and largest section, labeled \"(b) Vehicle-Infrastructure Cooperative System\", illustrates a complex data flow architecture enclosed in dashed borders. The upper part of this section, shaded in light yellow, depicts \"Data Transmission for Three Fusion Schemes,\" showing a progression from \"Raw Sensor Data\" through \"Model\" and \"Feature\" blocks to \"Intermediate-Level Data\" and finally \"Detection Outputs\". Below this, separated by a header reading \"Limited Communication Condition\", the lower part shows a street view with a crosswalk and four vehicles (white, blue, green, and red). Lines connect these vehicles to a transmission line with envelope icons, illustrating delays labeled \"After $\\Delta t_1$\", \"After $\\Delta t_2$\", \"After $\\Delta t_3$\", and \"After $\\Delta t_N$\". On the far right, section (c) displays two grayscale images showing overlapping sensor data with purple and red circles highlighting objects, captioned \"(c) Aligned Features / Point Clouds\". Directly beneath this, section (d) shows a similar pair of images where the highlighted objects are misaligned, captioned \"(d) Non-Aligned Features / Point Clouds\". The entire visual layout uses a palette of white, light blue, light yellow, and gray, with blue and black text to convey the hierarchical structure of the object detection challenges and system architecture.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/0cabf7c1-b7b7-4ce0-82c9-9d3c43ed9f88_input_0.jpg"], "target_image": "images/multi_image/0cabf7c1-b7b7-4ce0-82c9-9d3c43ed9f88_target.png", "width": 2592, "height": 1461}
+{"index": 99, "id": "f04449d3-a106-4e81-b8b5-b414c3a4d1ed", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, titled **GAN-based Model Inversion Attack** in a large, bold sans-serif font at the top, features a light beige background and is largely enclosed by a red dashed rectangular border. To the immediate left of the title, a decorative element consists of two overlapping squares, one outlined in blue and the other solid dark blue.\n\nThe upper-left quadrant of the slide outlines the context and objectives of the attack using black text. A bullet point states: \"First proposed by GMI[1]\". Below this, another bullet point reads \"White-box accessibility\", followed by three indented sub-points detailing the required access: \"Output information\", \"Parameters of target model\", and \"Auxiliary public dataset\". Further down in this section, a third main bullet point labeled \"Goal\" is followed by the specific objectives: \"Recover private data or features\" and \"Usefulness in downstream tasks\".\n\nOccupying the upper-right and central portion of the slide is a detailed flowchart illustrating the technical process. The diagram begins on the left with a green cloud shape containing the mathematical notation \"Z ~ N (0,I)\". An upward arrow labeled \"Sample\" points from the cloud to a group of small green blocks representing vectors, separated by ellipses, labeled \"Latent Vectors\". An arrow extends from these vectors to a cluster of vertical blue rectangular planes labeled \"Generator\". Output from the generator flows to a pair of portrait photographs labeled \"Generated Images\", separated by ellipses.\n\nFrom the generated images, the flow splits. One path leads right to a cluster of orange rectangular planes labeled \"Target Model\", which outputs to a blue box containing the text \"Target Labels\" and an arrow labeled \"Lidentity\". Another path from the generated images goes down to a cluster of green rectangular planes labeled \"Discriminator\", which outputs an arrow labeled \"Lprior\". The \"Lidentity\" and \"Lprior\" lines converge into a circle with a plus sign, summing up to an upward arrow labeled \"Ltotal\". A feedback loop arrow extends from the top \"Ltotal\" line all the way back to the \"Latent Vectors\", labeled \"Optimize\". Beneath this diagram, centered in small text, is the caption: \"Fig.2 An overview of GAN-based MIA\".\n\nSeparated by a horizontal dashed red line, the lower section of the slide depicts the operational flow under the heading \"Pipeline\". This visual is a horizontal chevron process flow consisting of four light brown arrow-shapes pointing right. The first arrow contains the text \"Pre-train the GAN\", the second says \"sample\", the third says \"generate\", and the fourth concludes with \"Optimize iteratively\".\n\nAt the very bottom of the slide, a footer in small grey text provides the date \"2024/10/4\" on the left, followed by the citation: \"[1] Zhang et. al. The secret revealer: Generative model-inversion attacks against deep neural networks. In CVPR 2020.\" On the far bottom right is the page number \"3\".\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/f04449d3-a106-4e81-b8b5-b414c3a4d1ed_input_0.jpg"], "target_image": "images/multi_image/f04449d3-a106-4e81-b8b5-b414c3a4d1ed_target.png", "width": 2880, "height": 1620}
+{"index": 100, "id": "1a308c06-6b42-4200-9e3b-67165e5f6a3c", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, set against a clean white background, is titled \"1) Experimental results with only stragglers\" in a large, dark blue sans-serif font at the top left. The main body of the slide features three distinct line charts arranged horizontally across the center, each representing a different dataset experiment. The first chart on the left corresponds to the **MNIST** dataset; its vertical axis is labeled \"Test accuracy\" ranging from 30 to 100, and the horizontal axis is labeled \"Running time\" from 0 to 80. This chart contains a legend identifying five methods: a blue line for \"Safeflow (Ours)\", an orange line for \"FedAsync\", a yellow line for \"Ignore stragglers + FedAvg\", a purple line for \"Wait for stragglers + FedAvg\", and a green line for \"Wait for partial stragglers + FedAvg\". The second chart in the center is for the **FMNIST** dataset, with a \"Test accuracy\" axis from 30 to 90 and a \"Running time\" axis from 0 to 100. It utilizes the same five-colored legend entries as the first chart. The third chart on the right represents the **CIFAR-10** dataset, with a \"Test accuracy\" axis ranging from 10 to 70 and a \"Running time\" axis extending from 0 to 1200. Its legend lists \"Sageflow (Ours)\" (blue), \"FedAsync\" (orange), \"Ignore stragglers + FedAvg\" (yellow), \"Wait for stragglers + FedAvg\" (purple), and \"Wait for partial stragglers + FedAvg\" (green). Below the charts, centered text reads \"MNIST\" under the first graph, \"FMNIST\" under the second, and \"CIFAR-10\" under the third. Beneath the charts, a caption in bold black text states \"Figure 1: Performance with only stragglers and contact ratio C=0.1\". In the bottom right corner of the slide, the page number \"- 26 -\" is visible in gray. The input includes sub-images of these charts, which occupy the central horizontal band of the slide: sub-image 1 is on the left, sub-image 2 is in the middle, and sub-image 3 is on the right, each taking up roughly a third of the slide's width. No other decorative elements, borders, or background textures are present.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/1a308c06-6b42-4200-9e3b-67165e5f6a3c_input_0.jpg", "images/multi_image/1a308c06-6b42-4200-9e3b-67165e5f6a3c_input_1.jpg", "images/multi_image/1a308c06-6b42-4200-9e3b-67165e5f6a3c_input_2.jpg"], "target_image": "images/multi_image/1a308c06-6b42-4200-9e3b-67165e5f6a3c_target.png", "width": 1728, "height": 1296}
+{"index": 101, "id": "4de22691-81c0-4273-885a-c856fdba5d39", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean white background with a prominent deep red banner at the top containing the title \"Optimal Transport (OT) -based PL\" in large, bold white text. In the top right corner, the red logo and name of \"ShanghaiTech University\" (上海科技大学) are displayed. The main visual focus is a central diagram that flows from left to right, illustrating a mathematical process. On the left side, under the label \"class centroids\" in yellow text, is a matrix labeled \"Prediction matrix P\". This matrix contains three columns titled \"bike\", \"backpack\", and \"glasses\" in yellow text, and five rows corresponding to icons of green bicycles, green backpacks, and green glasses, which are collectively labeled \"samples\" in green text. The matrix cells are shaded in varying intensities of blue, indicating probability distributions. An arrow points from this section to the right, with the text \"Given $\\alpha, \\beta$, solving OT\" above it. Below this text are visual representations of two vectors: a green bar chart labeled $\\alpha$ showing six equal-height bars, and a yellow bar chart labeled $\\beta$ showing three equal-height bars. The result on the right is a matrix labeled \"Coupling matrix Q*\" with the same row icons and column headers. This matrix shows a distinct pattern of dark blue diagonal blocks and white empty cells. To the right of this matrix is a green vertical bar chart labeled $\\alpha$, and below it is a yellow horizontal bar chart labeled $\\beta$. A pink rounded rectangle points to the resulting structure with the text \"Pseudo-labeling matrix\". At the bottom of the slide, a beige rectangular box contains the mathematical formulation titled \"Optimal Transport-based PL\". Inside this box is the minimization formula: \"$\\min_{\\mathbf{Q}\\in\\Pi(\\frac{1}{B}\\mathbb{I}_{B},\\frac{1}{C}\\mathbb{I}_{C})}\\langle-\\log\\mathbf{P},\\mathbf{Q}\\rangle+\\varepsilon\\langle\\mathbf{Q},\\log\\mathbf{Q}\\rangle$\". Directly below this is the set definition: \"$\\mathbf{\\Pi}(\\boldsymbol{\\alpha},\\boldsymbol{\\beta})=\\left\\{\\mathbf{Q}\\in\\mathbb{R}_{+}^{|\\alpha|\\times|\\beta|}\\middle|\\mathbf{Q}\\mathbb{I}_{|\\beta|}=\\boldsymbol{\\alpha},\\mathbf{Q}^{\\mathrm{T}}\\mathbb{I}_{|\\alpha|}=\\boldsymbol{\\beta}\\right\\}$\". The footer of the slide features a silhouette of buildings on the left and the Chinese motto \"立志成才报国裕民\" in the bottom right corner.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/4de22691-81c0-4273-885a-c856fdba5d39_input_0.jpg"], "target_image": "images/multi_image/4de22691-81c0-4273-885a-c856fdba5d39_target.png", "width": 1920, "height": 1080}
+{"index": 102, "id": "d0e1896c-dc78-4882-a535-eca7dcffb648", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide, titled **VALUES**, features a clean, high-contrast design with a black header section and a white main body. The header prominently displays the word \"VALUES\" in large, bold, white sans-serif capital letters centered at the top. Flanking the title on both the far left and far right are decorative, white, wireframe-style wave graphics that resemble flowing ribbons or topographic lines, adding a dynamic yet subtle texture to the black background.\n\nBelow the header, the slide is divided into a two-by-two grid layout containing two distinct sections: \"Mission\" and \"Vision.\" The upper-left quadrant is dedicated to the **Mission**. It begins with a sub-image labeled \"Mission\" at the top, showing a workspace scene with a laptop displaying a video conference grid, a coffee cup, and a smartphone being held by a person. Directly to the right of this image is a text block with a black border containing the following verbatim statement: \"• Our art residency program aims to provide a supportive multicultural environment for artists working at the intersection of traditional art, culture, and technology. We strive to foster creativity, innovation, and experimentation, while promoting cross-disciplinary collaboration and community engagement\". The text is black, left-aligned, and set against a white background.\n\nThe lower portion of the slide presents the **Vision** statement. The layout here is mirrored compared to the top section. On the bottom-left, there is a bordered text box containing the verbatim text: \"• Our vision is to be a leading art residency program that supports artists in exploring the potential of mixing traditional art with technology, experienced it by music to create new forms of artistic expression mixed with different cultures.\" To the right of this text block is a second sub-image labeled \"Vision.\" This image depicts a large, multifaceted gemstone or crystal splashing into a pool of water, creating ripples and droplets, symbolizing clarity and impact. The visual elements are arranged to create a balanced flow from the mission statement on the top right to the vision image on the bottom right, while the images anchor the left side. The overall color palette is strictly black, white, and the natural colors of the photographs, maintaining a professional and focused aesthetic.\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/d0e1896c-dc78-4882-a535-eca7dcffb648_input_0.jpg", "images/multi_image/d0e1896c-dc78-4882-a535-eca7dcffb648_input_1.jpg"], "target_image": "images/multi_image/d0e1896c-dc78-4882-a535-eca7dcffb648_target.jpg", "width": 2048, "height": 1152}
+{"index": 103, "id": "5cc09500-beb0-4572-a0fe-d50fcf216eb9", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, white background with a minimalist layout designed to highlight visual evidence of a workshop or meeting. At the very top left corner, a small logo is visible, composed of two overlapping triangles—one yellow and one teal—situated above a grid of blue dots. To the right of this logo, the title text \"PHOTO DOCUMENTATION\" is displayed in a large, bold, black sans-serif font.\n\nThe central portion of the slide is dominated by three photographic sub-images arranged horizontally. The first sub-image, positioned on the left, shows a group of people seated around a wooden table under a covered outdoor area; a person on the far left is holding up a phone, appearing to take a picture. The middle sub-image depicts a slightly closer view of participants seated on teal plastic chairs, with one woman in a green shirt using a laptop at a wooden table while others listen. The third sub-image, located on the right, presents a wider angle of the gathering, showing more attendees seated in the semi-open space with a computer monitor visible in the foreground. Beneath this row of three sub-images, the caption text \"Facilitation of Enterprise Assessment\" is centered on the slide in a smaller, grey sans-serif font, describing the activity shown in the photos.\n\nIn the bottom right corner of the slide, a decorative abstract element adds a splash of color to the footer area. This design consists of several geometric shapes, including a diagonal blue bar, a green parallelogram, and a yellow-orange stripe, angled upward to the right. The overall composition is balanced and professional, using the white space to draw focus to the photographic documentation and the accompanying text labels.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/5cc09500-beb0-4572-a0fe-d50fcf216eb9_input_0.jpg", "images/multi_image/5cc09500-beb0-4572-a0fe-d50fcf216eb9_input_1.jpg", "images/multi_image/5cc09500-beb0-4572-a0fe-d50fcf216eb9_input_2.jpg"], "target_image": "images/multi_image/5cc09500-beb0-4572-a0fe-d50fcf216eb9_target.jpg", "width": 2048, "height": 1152}
+{"index": 104, "id": "db93018b-b3c2-4362-8a70-31f075b118cf", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide is visually split into two distinct vertical sections: a photographic image on the left and a column of text on the right, set against a cream-colored background. The left section is entirely occupied by sub-image 1, which depicts the interior of a fortune-telling arcade booth. The image features an animatronic figure labeled \"ZOLTAR\" in large, arched red lettering at the top. The figure, dressed in purple and gold robes with a turban, sits behind a glass front. On the counter before the figure are a glowing crystal ball, tarot cards, and two decorative skulls—one plain and one studded with jewels. Warm orange lamps flank the figure's head against a red curtain backdrop.\n\nThe right side of the slide presents a critique of the \"Future of Jobs Report\" in black sans-serif text. The section begins with a headline that reads: \"The Future of Jobs Report relies heavily on a survey of large companies' executives. This presents several critical limitations.\" Within this introductory block, specific phrases are emphasized with a yellow highlight: \"The Future of Jobs Report\", \"companies' executives\", and \"presents several critical limitations\".\n\nBelow the introduction, the text is organized under two main subheadings. The first is titled \"Uncertainty and the 'Unknown Unknowns' of AI\". Following this header are three bullet points detailing specific issues. The first bullet states: \"Rapid Technological Change: The pace of AI development makes accurate mid- to long-term prediction extremely challenging.\" The second bullet reads: \"Limited Expertise: Survey respondents are not necessarily AI experts; they are making projections based on current (and potentially flawed) understanding.\" The third bullet explains: \"'Unknown Unknowns': The most disruptive impacts may be those we cannot currently foresee.\"\n\nThe second major section is titled \"Biases and the 'Ackoff Problem'\". It also contains three bullet points. The first reads: \"Executive Bias: The survey captures the perspective of large-company executives, not workers or smaller businesses.\" The second bullet states: \"Self-Reported Data: Responses may be subject to biases, organizational pressures, and the desire to appear forward-thinking.\" The final bullet point on the slide reads: \"Ritualistic Planning: The report may reflect aspirations rather than realistic assessments, echoing Ackoff's critique of corporate planning: '...like a ritual rain dance...'\" The layout relies on clear typography with bolded headers and bolded key terms (like \"Rapid Technological Change\" and \"Executive Bias\") to structure the information, while the contrast between the whimsical sub-image 1 and the serious analytical text creates the slide's primary visual metaphor.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/db93018b-b3c2-4362-8a70-31f075b118cf_input_0.jpg"], "target_image": "images/multi_image/db93018b-b3c2-4362-8a70-31f075b118cf_target.jpg", "width": 2048, "height": 1152}
+{"index": 105, "id": "76a5ef38-fd56-4e13-b705-6f6984493fef", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide employs a striking split-screen design, balancing a complex digital illustration on the left with concise informational text on the right. The left side of the slide is entirely filled by a sub-image, which occupies the full height and approximately the left third to half of the width. This sub-image depicts a side-profile portrait of a man with dark hair and a beard, set against a background featuring a large, textured orange circle and hints of grunge-style splatter. The right side of the man's head and neck is stylized as a dark, translucent digital grid, revealing layers of code, circuitry, and small floating screens or data windows embedded within his silhouette, symbolizing the integration of technology and humanity.\n\nThe right side of the slide features a clean, white background that provides high contrast for the black text content. At the very top right, serving as the slide's main heading, is the text \"History and Evolution of AI\" in a large, bold, sans-serif font. The text is broken into two lines, with \"History and Evolution of\" on the first line and \"AI\" emphasized on the second line. Below this title, occupying the majority of the right-hand space, is a single paragraph of body text in a standard serif font. The text begins with the sentence: \"The roots of Artificial Intelligence date back to the mid-20th century, with pioneers like Alan Turing and John McCarthy laying foundational concepts.\" Following immediately is the second sentence: \"Early AI focused on problem-solving and logic-based computation.\" The narrative continues with: \"The field experienced various phases, including the first AI winter periods due to unmet expectations followed by resurgence in the 21st century driven by advancements in machine learning, deep learning, and vast data availability.\" The paragraph concludes with the final sentence: \"Today, AI technology is integrated into everyday applications, continuously evolving through research and investment.\" The layout is minimalist on the text side, with generous margins ensuring readability, while the visual side adds a thematic, artistic layer of depth regarding the subject matter.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/76a5ef38-fd56-4e13-b705-6f6984493fef_input_0.jpg"], "target_image": "images/multi_image/76a5ef38-fd56-4e13-b705-6f6984493fef_target.jpg", "width": 2048, "height": 1152}
+{"index": 106, "id": "ebe562ef-c879-405d-8e30-2ab959d3c4ed", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a split-screen layout with a prominent photographic element on the left and a block of informational text on the right, all set against a clean, light grey background. Dominating the left side is a large, high-resolution photograph of a male lion with a thick, golden-brown mane standing amidst tall, dry savanna grass. The lion looks directly forward, and the background of the photo fades into a soft blue sky. A distinct, jagged white border separates this image from the text area on the right, giving the impression of torn paper.\n\nAt the top right, the slide's main title is displayed in a bold, dark sans-serif font, reading: \"Opportunistic vs. Equilibrium Species\". Immediately below the title, a secondary heading labeled \"EQUILIBRIUM SPECIES\" is highlighted in black uppercase letters, set against a bright yellow rectangular background to draw attention. The main body of text is organized into two distinct bullet points. The first bullet point explains the life-history characteristics of these animals, stating: \"They typically reach their reproductive age slower than opportunistic species, have longer lives, and have received extended parental care.\" This text is aligned to the left, just below the yellow banner. The second bullet point, positioned further down the slide, provides examples of such animals: \"Birds, large mammals, and species with type I or type II survivorship curves are often equilibrium species\". The overall color palette is natural and academic, utilizing the earth tones of the lion and grass, the bright yellow accent for emphasis, and standard dark grey text for readability. The layout effectively juxtaposes the visual example of a large mammal with the theoretical definition of an equilibrium species.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/ebe562ef-c879-405d-8e30-2ab959d3c4ed_input_0.jpg"], "target_image": "images/multi_image/ebe562ef-c879-405d-8e30-2ab959d3c4ed_target.jpg", "width": 2048, "height": 1152}
+{"index": 107, "id": "324f8cda-7e35-499c-958c-1cfc6390f015", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide utilizes a clean, high-contrast white background to emphasize its text-heavy content, arranged in a standard two-column layout. The left side is dominated by black text in a modern sans-serif font, while the right side features a prominent portrait of a man. The slide begins at the top left with a section header that poses the central question: \"Why is important to know our origin?\" immediately establishing the theme of the presentation. Below this heading, a substantial block of text presents a quote enclosed in quotation marks. This quote begins with the sentence \"Believing in intelligent design as opposed to chance evolution will have a profound impact on our self-concept and what we consider to be the ultimate purpose for our lives.\" It continues directly into the next sentence: \"Our sense of who we are is always connected with our understanding of where we came from.\" The final sentence of the quote reads: \"Not understanding our divine origin can have a catastrophic negative impact on our life and on the world\". Directly beneath this quote, without a paragraph break, the slide provides the source citation: \"Who do you think you are? Dough Batchelor, p.11\".\n\nTo the right of the text, occupying roughly the right third of the slide's width and aligned vertically with the text block, is sub-image 1. This image depicts a man with a bald head and a gray mustache, wearing wire-rimmed glasses. He is dressed formally in a navy blue suit jacket, a crisp white dress shirt, and a tie featuring diagonal stripes in alternating shades of blue, silver, and white. He is smiling warmly at the camera and holding a black book with gold lettering on the spine that reads \"HOLY BIBLE\" against a soft-focus background of purple and blue horizontal gradients. The overall composition balances the heavy textual argument on the left with the authoritative, personable visual presence of the speaker or author on the right, using ample negative space to prevent the slide from feeling cluttered. There are no decorative icons, borders, or footer elements visible on the slide; the focus remains strictly on the textual message and the accompanying portrait.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/324f8cda-7e35-499c-958c-1cfc6390f015_input_0.jpg"], "target_image": "images/multi_image/324f8cda-7e35-499c-958c-1cfc6390f015_target.jpg", "width": 2048, "height": 1152}
+{"index": 108, "id": "36bcf477-7626-40ed-ad74-121da4b81e79", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide is set against a background that gradients smoothly from a deep moss green on the left side to a calming slate blue on the right. Faint, concentric white arcs curve across the upper right corner, adding a subtle texture to the blue area, while similar curved lines appear at the bottom left in a lighter green hue. In the top right corner, a white speech bubble logo contains the OECD emblem followed by the text \"Green Talks: LIVE\".\n\nOn the left side of the slide, a vertical white card with rounded corners displays the OECD logo at the top, followed by the title text in dark blue font: \"Transforming Catalonia’s Mobility System for Net Zero\". Below this header section, sub-image 1 occupies the lower portion of the card. Directly beneath this entire white card element, the URL \"https://oe.cd/cc-catalonia\" is printed in large, white sans-serif letters, centered horizontally relative to the card above it.\n\nIn the center of the slide sits sub-image 2, which acts as a standalone visual element slightly offset to the right of the first card. It features a green rectangular banner near the top right with the white text \"POLICY HIGHLIGHTS\". At the bottom right corner of this sub-image, the OECD logo appears again, accompanied by the tagline \"BETTER POLICIES FOR BETTER LIVES\".\n\nTo the far right of the layout, sub-image 3 is positioned within a dark green rounded square frame. Below this frame, the instruction \"Scan me\" is written in white text. The bottom right corner of the slide features the OECD logo and name in white, grounding the overall composition. The visual arrangement guides the eye from the informational card and link on the left, through the central imagery, to the interactive QR code element on the right.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/36bcf477-7626-40ed-ad74-121da4b81e79_input_0.jpg", "images/multi_image/36bcf477-7626-40ed-ad74-121da4b81e79_input_1.jpg", "images/multi_image/36bcf477-7626-40ed-ad74-121da4b81e79_input_2.jpg"], "target_image": "images/multi_image/36bcf477-7626-40ed-ad74-121da4b81e79_target.jpg", "width": 2048, "height": 1152}
+{"index": 109, "id": "16ae8828-0ee5-4120-960a-64266fa2d91a", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean white background and is designed to visually explain the limitations of IPv4. At the top, a large, bold blue headline asks the central question: \"Why can't we keep using IPv4?\" Centered directly beneath this title is a square graphic element—referenced as sub-image 1—which serves as a visual metaphor.\n\nPositioned immediately below this sub-image is a concise black sentence: \"We've run out!\" The layout then transitions to a more technical explanation at the bottom of the slide. The first line of this explanatory text reads, \"IPv4 uses a 32-bit address for Internet addresses.\" Following this, the next sentence elaborates on the mathematical limit, stating, \"It can provide support for 2^32 IP addresses in total around 4.29 billion.\" Notably, the number \"4.29 billion\" in this final sentence is highlighted in red text to draw attention to the specific magnitude of the address space.\n\nAt the very bottom left of the slide, the acronym \"APNIC\" appears in black. Spanning horizontally across the bottom is a thin blue line that acts as a footer divider. In the bottom right corner, the slide number \"18\" is visible alongside a faint logo consisting of circular graphical elements. The overall design relies on high contrast with blue headers and red accents to emphasize the critical nature of the IP address depletion issue.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/16ae8828-0ee5-4120-960a-64266fa2d91a_input_0.jpg"], "target_image": "images/multi_image/16ae8828-0ee5-4120-960a-64266fa2d91a_target.jpg", "width": 2048, "height": 1152}
+{"index": 110, "id": "8b21c39c-7431-48f1-96e3-f92a130cfec6", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation title slide is designed with a clean, professional split-screen layout, featuring a vertical photograph on the left side and text content on a white background to the right. On the left, which occupies roughly one-third of the slide's width, is a full-height sub-image (sub-image 1) depicting two women engaged in a professional discussion within an office setting.\n\nTo the right of this photograph, the slide's main heading is displayed in a large, bold, blue sans-serif font across three lines. The text reads, \"Ensuring a Respectful Workplace: The POSH Act in India.\"\n\nBeneath this title, a paragraph of introductory text appears in a smaller, standard black font. This text begins by stating, \"This presentation offers an in-depth look into the Prevention of Sexual Harassment of Women at Workplace (Prevention, Prohibition and Redressal) Act, 2013, commonly known as the POSH Act.\" Immediately following this, the paragraph continues, \"We will explore its foundational principles, scope, and the critical mechanisms for ensuring a safe and dignified environment for all employees in India.\" The background of the text section is a plain, solid white, providing high contrast for the blue title and black body text, ensuring legibility. No other graphical elements, icons, or decorations are visible on the slide.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/8b21c39c-7431-48f1-96e3-f92a130cfec6_input_0.jpg"], "target_image": "images/multi_image/8b21c39c-7431-48f1-96e3-f92a130cfec6_target.jpg", "width": 2048, "height": 1152}
+{"index": 111, "id": "1733537b-3684-495f-8489-7cf3d5ea8861", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "Set against a solid, muted teal-blue background, this presentation slide employs a clean, professional layout that balances typography on the left with abstract geometric artistry on the right. Dominating the upper-left quadrant is a headline in a large, dark navy sans-serif font that declares: \"In a connected and mobile world, talent is the engine of economic growth and the drive toward a knowledge-based society.\" Situated directly beneath this headline, a secondary block of text appears in a smaller, lighter gray font. This paragraph elaborates on the central theme, stating: \"Most, if not all, organizations are increasingly aware that a knowledge-based society depends on the best talent being in place across key sectors of the economy. Countries that have established a strong, talent-based workforce – through attracting the best from overseas – have developed thriving, sustainable economies.\"\n\nTo the right of the text, occupying roughly the right third of the slide, is a sophisticated graphical arrangement composed of hexagonal shapes. This cluster resembles a molecular structure or a honeycomb network, featuring hexagons in varying sizes and states of opacity. The hexagons display a range of colors, including deep blues, sky blues, and lighter grayish-blues, alongside several hexagons that are rendered as white outlines with transparent centers. These shapes are interconnected by thin white lines that trace the edges of the grid, with small, dark blue dots marking specific intersection points, adding a sense of connectivity and structure to the design.\n\nIn the top-right corner, the slide number \"/03\" is discreetly printed in a small, white font. Anchoring the bottom of the slide is a simple white horizontal line that extends across most of the width, stopping just before the right edge, serving as a footer element that visually grounds the content. The overall design aesthetic is modern and corporate, utilizing a monochromatic color palette to reinforce themes of connection and structure.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/1733537b-3684-495f-8489-7cf3d5ea8861_input_0.jpg"], "target_image": "images/multi_image/1733537b-3684-495f-8489-7cf3d5ea8861_target.jpg", "width": 2048, "height": 1331}
+{"index": 112, "id": "d838a461-f108-414f-904f-142028a3747f", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a striking educational layout divided into two distinct sections. The upper two-thirds of the slide is dominated by a large, high-resolution photograph of a hammerhead shark swimming underwater. The shark, positioned diagonally with its head near the top left and tail extending toward the middle right, displays a sleek, gray body with a contrasting white underbelly. Sunlight filters down from the ocean surface visible at the top of the frame, illuminating the shark's dorsal fin and highlighting the distinctive, wide, hammer-shaped head known as a cephalofoil. A small remora fish is visible clinging to the shark's side near the tail. This photograph serves as a full-width header image, referred to as sub-image 1, sitting atop a white, lower section that contains the textual information.\n\nIn the lower white section, the text is organized into two columns. On the left side, the main title appears in a large, bold, black sans-serif font, reading: \"Hammerhead Shark (Sphyrnidae family)\". To the right of the title, a bulleted list provides detailed facts about the species, written in a smaller, standard-weight black font. The list begins with the size description: \"Size: Varies by species; Great Hammerhead can reach 20 feet (6 meters), while smaller species are around 10–13 feet (3–4 meters).\" The next point details the shark's food sources: \"Diet: Primarily feeds on fish, squid, stingrays, and crustaceans.\" Following this is information on where the shark lives: \"Habitat: Warm coastal waters and continental shelves worldwide.\" The behavior of the shark is then described: \"Behavior: Hunts in schools during the day and becomes solitary at night.\" The final bullet point highlights the shark's anatomy and sensory capabilities: \"Unique Feature: Wide, hammer-shaped head (“cephalofoil”) enhances vision and electroreception for detecting prey like stingrays buried in sand.\" The overall design uses a clean white background for the text area to ensure readability, with a smooth, curved bottom edge on the image separating it from the text content below.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/d838a461-f108-414f-904f-142028a3747f_input_0.jpg"], "target_image": "images/multi_image/d838a461-f108-414f-904f-142028a3747f_target.jpg", "width": 2048, "height": 1152}
+{"index": 113, "id": "d7bebc76-158c-49a3-8bfb-83ce3aba809b", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean white background with a minimalist layout centered around a vertical bar chart and supporting text. At the very top, a large, dark grey heading in a sans-serif font reads \"2. Support professionals prioritise customer happiness above revenue\". Directly beneath this title, the central visual element is a bar chart displayed within the slide's layout (referenced as sub-image 1 in the provided list). This chart visualizes data regarding the prioritization of customer interactions. The vertical axis is labeled \"Number of responses\" and includes numerical markers for 0, 10, 20, 30, and 40. The horizontal axis is labeled \"Handling customer enquiries\" and displays five distinct categories: \"Not a priority\", \"2\", \"3\", \"4\", and \"Very important\".\n\nThe data is represented by coral-red vertical bars of varying heights. Corresponding to the \"Not a priority\" label on the x-axis, there is no visible bar, indicating a value of zero. Above the label \"2\", a very short bar sits just above the zero line. Above the label \"3\", a slightly taller bar reaches a height visually estimated to be around 2. Above the label \"4\", a significantly taller bar extends to the grid line corresponding to 10. The final category, \"Very important\", features the tallest bar by a large margin, extending well past the 30 mark to approximately 35 on the vertical scale. Positioned directly underneath the x-axis labels of the chart is a small, light grey sentence that states, \"This graph shows the number of responses for degree of priority across all roles.\"\n\nBelow the chart area, a large, bold statement in dark grey text serves as a summary of the data: \"95% of support professionals say handling customer enquiries is top priority\". This text is centered on the slide. The design concludes with a solid dark purple footer bar spanning the entire bottom width of the slide. Inside this footer, on the left side, the brand name \"kayako\" appears in white lowercase letters. On the far right side of the footer, the hashtag \"#custserv\" is displayed in white text. The overall aesthetic is professional and data-driven, utilizing a limited color palette of white, dark grey, coral red, and dark purple.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/d7bebc76-158c-49a3-8bfb-83ce3aba809b_input_0.jpg"], "target_image": "images/multi_image/d7bebc76-158c-49a3-8bfb-83ce3aba809b_target.jpg", "width": 2048, "height": 1536}
+{"index": 114, "id": "d3749782-1e2f-4cb6-81c2-d8c8a546c868", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide is titled with the bold, black question \"What makes Zoho Cloud Email Secure?\" centered at the top against a clean white background. Dominating the left half of the slide is a large photograph of an iceberg floating in the ocean, serving as a visual metaphor for security layers. The image depicts a clear horizon line where the sky meets the water; the visible portion of the iceberg is jagged and snow-covered, while a much larger, massive structure extends deep into the translucent blue water below. Two yellow callout bubbles labeled \"1\" and \"2\" point to the surface tip and the submerged base respectively.\n\nTo the right of the iceberg image, the slide presents a two-point list that explains the metaphor. The first point is headed with the bold blue text \"1. Tip of Iceberg.\" Immediately following this heading is the explanatory sentence: \"Security features users can see and manage,\" written in a standard black sans-serif font. Further down the slide, aligned beneath the first point, is the second heading in blue: \"2. The base of the Iceberg.\" This section is followed by the description: \"Foundational Security that lies underneath - DC & Architecture level,\" which appears in black text. In the bottom right corner of the slide, a small, colorful Zoho logo is visible, consisting of geometric shapes in red, green, blue, and orange.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/d3749782-1e2f-4cb6-81c2-d8c8a546c868_input_0.jpg"], "target_image": "images/multi_image/d3749782-1e2f-4cb6-81c2-d8c8a546c868_target.jpg", "width": 2048, "height": 1280}
+{"index": 115, "id": "40dd4ee9-5b29-4fb7-9b57-489e0d9153c5", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, high-contrast design split into two distinct sections: a visual illustration on the left and a detailed textual analysis on the right. The overall color palette is strictly grayscale and black, conveying a serious and analytical tone. The background of the entire slide is a solid, deep black on the right, fading into a light gray backdrop on the left side behind the illustration.\n\nThe left third of the slide is dominated by a large, vertically oriented line-art illustration, identified as sub-image 1, which occupies roughly 35% of the total slide width. This graphic depicts a traditional balance scale. The left pan of the scale is heavily weighed down and filled with a chaotic pile of icons representing social media activity, including heart shapes, mobile phones, peace signs, and profile avatars. In contrast, the right pan is elevated and holds only a single, solitary padlock, symbolizing security or privacy. The scale is tipping towards the social media icons, visually suggesting that the weight of digital engagement is currently overwhelming the weight of security.\n\nOn the right side, against the black background, the content is presented in crisp white text. At the top, the main title reads \"Navigating Digital Challenges\" in a large, bold sans-serif font. Below this, the slide organizes key issues into three rectangular boxes with thin white borders. The top-left box is titled \"Misinformation Crisis\" and contains the text: \"Fake news spreads six times faster than truth on social platforms, requiring urgent solutions for information verification and fact-checking systems.\" To its right, the top-right box is labeled \"Privacy Concerns\" and states: \"Data breaches and unauthorized sharing of personal information threaten user trust and require stronger platform accountability and regulation.\" Below these two, occupying the full width of the text column, is a third box titled \"Digital Wellbeing,\" which reads: \"Cyberbullying, addiction, and mental health impacts demand media literacy education and healthier platform design focused on user welfare.\"\n\nBeneath these boxed sections, a concluding sentence spans the width of the text area, stating: \"Addressing these challenges requires collaboration between users, platforms, governments, and educators to harness social media's benefits while minimizing harm.\" Finally, in the bottom-right corner, a small, rounded rectangular button displays the text \"Made with GAMMA\" in blue and white.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/40dd4ee9-5b29-4fb7-9b57-489e0d9153c5_input_0.jpg"], "target_image": "images/multi_image/40dd4ee9-5b29-4fb7-9b57-489e0d9153c5_target.jpg", "width": 2048, "height": 1432}
+{"index": 116, "id": "017ea51e-db4c-48d9-9524-2a2b74955532", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a split-screen layout with a dark blue, textured background on the left two-thirds and a full-bleed photograph on the right vertical third. The overall design conveys a formal sports educational theme. On the left side, set against a deep blue backdrop that features a faint, patterned texture resembling a chain-link fence, the main title \"RULES IN SOFTBALL\" is displayed at the top left in large, bold, italicized white capital letters. Immediately to the right of the title, a stylized icon of a softball with glowing white stitching and a neon blue outline adds a visual accent. Below the title and slightly indented to the left, a solid light blue rectangular tag contains the section header \"Pitching\" in white sans-serif text.\n\nThe body of the slide lists two specific regulations in white text, aligned to the left. The first point begins with the number \"1.\" followed by the verbatim text: \"Both feet of the pitcher should be on the ground and he or she is allowed only to step one foot forward before pitching.\" Below this, the second point is numbered \"2.\" and reads: \"Underhand pitching/ throwing is used. The ball should be held with both hands at the start of the pitch.\" The typography is clean and legible, contrasting sharply with the background.\n\nThe right side of the slide is occupied by a photograph that serves as sub-image 1, which extends from the top edge to the bottom edge of the frame. This image depicts a female softball player in a crouched stance on a dirt field, wearing a dark helmet, a white jersey with dark green sleeves, black pants, white socks, and white cleats. Her posture suggests she is ready for action, visually reinforcing the topic of the rules. At the very bottom right corner of the slide, overlaying the lower part of the photograph, is the attribution text \"Created by: Jumark M. Diezmo\" in white lettering. Small decorative elements, such as a sliver of a bright blue vertical bar separating the text and image areas, and faint green shapes at the very top corners, provide structural definition to the slide's composition.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/017ea51e-db4c-48d9-9524-2a2b74955532_input_0.jpg"], "target_image": "images/multi_image/017ea51e-db4c-48d9-9524-2a2b74955532_target.jpg", "width": 2048, "height": 1152}
+{"index": 117, "id": "1323d62a-91e7-4b74-a688-6a2b8750c4f5", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a dramatic split layout with a large photograph occupying the left half and a solid black background hosting white typography on the right half. The overall color palette is dominated by deep midnight blues, stark whites, and the rich yellow and navy tones visible within the photograph. The right side of the slide serves as the primary information carrier, beginning with the main title at the top: \"The History of American College Sports,\" rendered in a large, bold, sans-serif typeface. Immediately below this is a secondary, slightly larger title that reads \"Tradition, Triumph, and Transformation,\" emphasizing the key themes of the presentation. Beneath these headlines, a descriptive paragraph in a smaller, lighter-weight font provides context: \"From humble rowing competitions to billion-dollar television contracts, American college athletics have evolved into a cultural phenomenon that captivates millions and shapes the fabric of higher education.\" At the very bottom right, aligned with the text block, is the website address \"dbbetbd.org\" displayed in a small, underlined font.\n\nThe left side of the slide is dedicated to a vivid, full-height sub-image labeled sub-image 1, which depicts a bustling American football stadium scene at night. The photo captures the energy of a packed crowd cheering from the stands, many of whom are waving large flags in navy blue and yellow. Bright stadium floodlights illuminate the scene from the upper left, casting beams through a hazy atmosphere, likely smoke or fog. Prominently displayed on the side of the stadium structure in the background is a large, illuminated sign reading \"VICTORY FIELD.\" In the foreground of the image, near the bottom left, a referee in a traditional black-and-white striped uniform stands on the green turf, facing away from the camera toward the crowd, anchoring the scene in the context of an active sporting event. The visual elements of the stadium lighting and the enthusiastic crowd visually reinforce the text's themes of triumph and cultural significance.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/1323d62a-91e7-4b74-a688-6a2b8750c4f5_input_0.jpg"], "target_image": "images/multi_image/1323d62a-91e7-4b74-a688-6a2b8750c4f5_target.jpg", "width": 2048, "height": 1170}
+{"index": 118, "id": "ac87eb21-7e54-41d8-b818-b0bed997861c", "task": "multi_image", "source": "real", "prompt_type": "long", "instruction": "This presentation slide features a clean, minimalist layout with a stark white background that draws attention to the content arranged in two distinct vertical sections. Dominating the right side is the slide’s primary textual message, anchored by the large, bold, black header \"What’s Excluded.\" Situated directly beneath this title is a vertically aligned bulleted list in a smaller, dark gray sans-serif font, which itemizes the specific costs not covered by the package. The list begins with the bullet point \"Tipping (driver/guide: ~USD 30–40/day; cook: ~USD 15–20/day),\" clearly delineating the expected gratuity amounts. Following this is the second item, \"Drinks (unless stated),\" indicating conditional coverage. The list continues with \"Visas & vaccinations\" on the third line, and concludes with \"Spending money & Zanzibar activities\" at the bottom.\n\nOccupying the left third of the slide is a rectangular photograph, referenced here as sub-image 1, which captures a wild scene. This sub-image shows a lioness lounging on a massive, weathered tree trunk in a dry, grassy savanna environment. The visual composition places the detailed text information on the right to balance the organic, textured imagery on the left. To add a subtle design touch to the lower-left corner, just below the photograph, there is a decorative graphic element consisting of four solid dark green triangles. These shapes are arranged horizontally in a row, pointing to the right, resembling a \"fast forward\" or \"progress\" icon, reinforcing the safari theme through color coordination with the vegetation in the photograph.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/ac87eb21-7e54-41d8-b818-b0bed997861c_input_0.jpg"], "target_image": "images/multi_image/ac87eb21-7e54-41d8-b818-b0bed997861c_target.jpg", "width": 2048, "height": 1152}
+{"index": 119, "id": "d589b5c6-1a48-4767-ae6e-020e0bd48c44", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#80B918\",\n      \"secondary\": \"#0055A4\",\n      \"text\": \"#000000\",\n      \"accent\": \"#8B0000\"\n    },\n    \"special_instructions\": \"The slide features a clean, high-contrast academic style. It utilizes a specific color coding system: Red for emphasis and conflict, Green for 'Sparse training' concepts, and Blue for 'Layer freezing' concepts. A large, semi-transparent grey watermark of the university logo is visible in the background.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Contrasting\",\n    \"composition_style\": \"Two-column comparison layout with a header and footer\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 4,\n      \"blocks\": [\n        {\n          \"id\": \"title-4\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Potential Issue of Incorporating Layer Freezing in DST\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Text is preceded by a solid red rectangular block on the far left.\"\n          }\n        },\n        {\n          \"id\": \"body-4-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Characteristics of sparse training may NOT compatible with layer freezing!\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The words 'may NOT compatible' are highlighted in red (#8B0000).\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A conceptual comparison diagram consisting of two columns. The left column is labeled 'Sparse training' and the right 'Layer freezing'. Two rows of comparison boxes are connected by red double-headed arrows labeled 'Feasibility?' and 'Potentiality?'.\",\n          \"style\": {\n            \"illustration_style\": \"Conceptual Diagram\",\n            \"color_scheme\": \"Green, Blue, Red, Black\",\n            \"position\": \"center\",\n            \"special_instructions\": \"Top row: Green box 'Keep mutating structure' vs Blue box 'Not allow to change'. Bottom row: Green box 'Sparsity may affect convergence speed' vs Blue box 'Benefit more from fast convergence'. Red arrows connect the rows.\"\n          }\n        },\n        {\n          \"id\": \"body-4-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"1. Sparse training need to keep mutating sparse structure, while layer freezing does not allow this mutation. Seems inherently contradict to each other.\\n\\n2. The impact of sparsity on layer learning speed.\\nLearning faster -> freeze earlier -> save more\\nLearning slower -> freeze later -> save less\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Numbered list. The word 'inherently' is bolded.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large, semi-transparent grey watermark of a university logo (Northeastern University shield and letters) placed in the background behind the central diagram.\",\n          \"style\": {\n            \"illustration_style\": \"Watermark\",\n            \"color_scheme\": \"Light Grey\",\n            \"position\": \"center-right background\",\n            \"special_instructions\": \"Low opacity, serves as a background element.\"\n          }\n        },\n        {\n          \"id\": \"body-4-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Northeastern University\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#8B0000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Located next to the university logo icon.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/d589b5c6-1a48-4767-ae6e-020e0bd48c44_input_0.jpg"], "target_image": "images/multi_image/d589b5c6-1a48-4767-ae6e-020e0bd48c44_target.png", "width": 2880, "height": 1620}
+{"index": 120, "id": "03f3e58a-978b-4401-90be-31bb8735f84c", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic/Technical Minimalist\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2B579A\",\n      \"secondary\": \"#D3D3D3\",\n      \"text\": \"#000000\",\n      \"accent\": \"#2B579A\"\n    },\n    \"special_instructions\": \"Clean, white background with high contrast. Uses a standard sans-serif font family (likely Arial). The visual style is strictly functional and academic, focusing on clarity of data visualization. The charts use a specific shade of medium blue for data points and a light grey for the diagonal reference line.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Grid-based, Three-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 8,\n      \"blocks\": [\n        {\n          \"id\": \"title-8\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Metrics\",\n          \"style\": {\n            \"color\": \"#2B579A\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Title is aligned to the left margin, positioned near the top edge. Color matches the primary accent blue.\"\n          }\n        },\n        {\n          \"id\": \"body-8-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Low error\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"above the first chart (left column)\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Label for the first chart. Centered horizontally above the chart.\"\n          }\n        },\n        {\n          \"id\": \"body-8-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"High rank correlation\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"above the second chart (center column)\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Label for the second chart. Centered horizontally above the chart.\"\n          }\n        },\n        {\n          \"id\": \"body-8-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Low regret\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"above the third chart (right column)\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Label for the third chart. Centered horizontally above the chart.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_174_423_907_1131.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_174_423_907_1131.jpg\"\n              }\n            ],\n            \"source\": \"Scatter plot illustrating 'Low error' with points on a diagonal line\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left column, below 'Low error' label\",\n            \"image_style\": \"Standard scatter plot with black axes, light grey diagonal line, and medium blue circular data points.\",\n            \"container\": null,\n            \"special_instructions\": \"The chart includes axis labels 'Estimate' (y-axis) and 'Value' (x-axis). The data points lie perfectly on the diagonal line.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1030_423_1769_1128.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1030_423_1769_1128.jpg\"\n              }\n            ],\n            \"source\": \"Scatter plot illustrating 'High rank correlation' with points above the diagonal line\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center column, below 'High rank correlation' label\",\n            \"image_style\": \"Standard scatter plot with black axes, light grey diagonal line, and medium blue circular data points.\",\n            \"container\": null,\n            \"special_instructions\": \"The chart includes axis labels 'Estimate' (y-axis) and 'Value' (x-axis). The data points are arranged in a line parallel to and above the diagonal reference line.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1887_404_2737_1128.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1887_404_2737_1128.jpg\"\n              }\n            ],\n            \"source\": \"Scatter plot illustrating 'Low regret' with points scattered and two specific points labeled\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, below 'Low regret' label\",\n            \"image_style\": \"Standard scatter plot with black axes, light grey diagonal line, and medium blue circular data points.\",\n            \"container\": null,\n            \"special_instructions\": \"The chart includes axis labels 'Estimate' (y-axis) and 'Value' (x-axis). The data points are scattered. Two specific points are labeled with text: 'Selected policy' (top right) and 'Best policy' (bottom right).\"\n          }\n        },\n        {\n          \"id\": \"illustration-8-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Three small logos stacked vertically in the bottom right corner\",\n          \"style\": {\n            \"illustration_style\": \"Corporate logos\",\n            \"color_scheme\": \"Blue, grey, and multi-colored dots\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Top logo is 'BAIR' with a bear icon. Middle logo is a grid of colored dots. Bottom logo is a blue swirl.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/03f3e58a-978b-4401-90be-31bb8735f84c_input_0.jpg", "images/multi_image/03f3e58a-978b-4401-90be-31bb8735f84c_input_1.jpg", "images/multi_image/03f3e58a-978b-4401-90be-31bb8735f84c_input_2.jpg"], "target_image": "images/multi_image/03f3e58a-978b-4401-90be-31bb8735f84c_target.png", "width": 2880, "height": 1620}
+{"index": 121, "id": "e9c752e5-a3e8-4f26-9b53-de180a2df616", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Conference Title Slide\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#0078D7\",\n      \"secondary\": \"#0078D7\",\n      \"text\": \"#000000\",\n      \"accent\": \"#0078D7\"\n    },\n    \"special_instructions\": \"Clean, minimalist academic presentation style. High contrast with a white background. The primary accent color is a specific shade of blue (#0078D7) used for the main title and the CVPR logo. Logos are placed in the top corners.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Academic, Technical\",\n    \"composition_style\": \"Centered\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 1,\n      \"blocks\": [\n        {\n          \"id\": \"logo-nvidia\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"The NVIDIA logo (green eye icon and black text) located in the top-left corner.\",\n          \"style\": {\n            \"illustration_style\": \"Logo\",\n            \"color_scheme\": \"Green and Black\",\n            \"position\": \"Top-left corner\",\n            \"special_instructions\": \"Standard NVIDIA branding.\"\n          }\n        },\n        {\n          \"id\": \"logo-ucla\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"The University of California, Los Angeles (UCLA) seal/logo located in the top-center-left area.\",\n          \"style\": {\n            \"illustration_style\": \"Logo\",\n            \"color_scheme\": \"Blue and White\",\n            \"position\": \"Top-center-left\",\n            \"special_instructions\": \"Standard UCLA branding.\"\n          }\n        },\n        {\n          \"id\": \"logo-stanford\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"The Stanford University seal/logo located in the top-center-right area.\",\n          \"style\": {\n            \"illustration_style\": \"Logo\",\n            \"color_scheme\": \"Red and White\",\n            \"position\": \"Top-center-right\",\n            \"special_instructions\": \"Standard Stanford branding.\"\n          }\n        },\n        {\n          \"id\": \"logo-cvpr\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"The CVPR (Conference on Computer Vision and Pattern Recognition) logo with date and location text.\",\n          \"style\": {\n            \"illustration_style\": \"Logo\",\n            \"color_scheme\": \"Blue and White\",\n            \"position\": \"Top-right corner\",\n            \"special_instructions\": \"Includes text 'JUNE 18-22, 2023' and 'VANCOUVER, CANADA'.\"\n          }\n        },\n        {\n          \"id\": \"title-1\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"FreeNeRF: Improving Few-shot Neural Rendering with Free Frequency Regularization\",\n          \"style\": {\n            \"color\": \"#0078D7\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"Top-center, vertically centered in the upper half\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold font. The text is split into two lines.\"\n          }\n        },\n        {\n          \"id\": \"body-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Jiawei Yang\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"Below the first author's photo, left side\",\n            \"container\": null,\n            \"special_instructions\": \"Standard sans-serif font.\"\n          }\n        },\n        {\n          \"id\": \"body-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"UCLA\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"Below 'Jiawei Yang'\",\n            \"container\": null,\n            \"special_instructions\": \"Standard sans-serif font.\"\n          }\n        },\n        {\n          \"id\": \"body-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Marco Pavone\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"Below the second author's photo, center\",\n            \"container\": null,\n            \"special_instructions\": \"Standard sans-serif font.\"\n          }\n        },\n        {\n          \"id\": \"body-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Nvidia Research, Stanford University\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"Below 'Marco Pavone'\",\n            \"container\": null,\n            \"special_instructions\": \"Standard sans-serif font.\"\n          }\n        },\n        {\n          \"id\": \"body-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Yue Wang\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"Below the third author's photo, right side\",\n            \"container\": null,\n            \"special_instructions\": \"Standard sans-serif font.\"\n          }\n        },\n        {\n          \"id\": \"body-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Nvidia Research\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"Below 'Yue Wang'\",\n            \"container\": null,\n            \"special_instructions\": \"Standard sans-serif font.\"\n          }\n        },\n        {\n          \"id\": \"image-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_688_991_1104_1405.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_688_991_1104_1405.jpg\"\n              }\n            ],\n            \"source\": \"Portrait of the first author, Jiawei Yang\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"Left side, vertically centered\",\n            \"image_style\": \"Circular crop\",\n            \"container\": null,\n            \"special_instructions\": \"The image is cropped into a perfect circle.\"\n          }\n        },\n        {\n          \"id\": \"image-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1656_992_2030_1405.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1656_992_2030_1405.jpg\"\n              }\n            ],\n            \"source\": \"Portrait of the second author, Marco Pavone\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"Center, vertically centered\",\n            \"image_style\": \"Circular crop\",\n            \"container\": null,\n            \"special_instructions\": \"The image is cropped into a perfect circle.\"\n          }\n        },\n        {\n          \"id\": \"image-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_2553_993_2964_1401.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_2553_993_2964_1401.jpg\"\n              }\n            ],\n            \"source\": \"Portrait of the third author, Yue Wang\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"Right side, vertically centered\",\n            \"image_style\": \"Circular crop\",\n            \"container\": null,\n            \"special_instructions\": \"The image is cropped into a perfect circle.\"\n          }\n        },\n        {\n          \"id\": \"body-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Paper tag: WED-AM-003\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"Bottom-left corner\",\n            \"container\": null,\n            \"special_instructions\": \"Small text indicating the paper session tag.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/e9c752e5-a3e8-4f26-9b53-de180a2df616_input_0.jpg", "images/multi_image/e9c752e5-a3e8-4f26-9b53-de180a2df616_input_1.jpg", "images/multi_image/e9c752e5-a3e8-4f26-9b53-de180a2df616_input_2.jpg"], "target_image": "images/multi_image/e9c752e5-a3e8-4f26-9b53-de180a2df616_target.png", "width": 3840, "height": 2160}
+{"index": 122, "id": "58cd9988-f0f7-46c2-8e88-50af0ef33ddc", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Samsung SDS Technical Results\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2B579A\",\n      \"secondary\": \"#D4A017\",\n      \"text\": \"#2B2B2B\",\n      \"accent\": \"#2B579A\"\n    },\n    \"special_instructions\": \"Clean, high-contrast technical presentation style. The slide features a stark white background with dark grey text. The charts utilize a standard grid system with light grey lines. The visual hierarchy is established through the large, bold title and the distinct separation of the two chart columns. The charts themselves are dense with data, using a variety of colors (blue, orange, red, green) to distinguish between different acquisition methods.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Data-Driven\",\n    \"composition_style\": \"Two-column grid layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 41,\n      \"blocks\": [\n        {\n          \"id\": \"title-41\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Results\",\n          \"style\": {\n            \"color\": \"#2B2B2B\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold sans-serif font. Positioned in the top-left corner of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-41-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"✓ 3 x CIFAR-100\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2B2B2B\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left, below the first chart\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Text is preceded by a checkmark icon. Positioned directly below the left chart.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_117_226_917_867.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_117_226_917_867.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing Accuracy vs Acquired dataset size for CIFAR-100\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left-center\",\n            \"image_style\": \"Standard chart with grid lines, legend, and axis labels\",\n            \"container\": null,\n            \"special_instructions\": \"The chart displays multiple line series with different markers and line styles. The legend is located in the bottom-right of the chart area. The x-axis is labeled 'Acquired dataset size' and the y-axis 'Accuracy'. A dotted horizontal line is visible near the top of the chart.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1005_227_1780_929.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1005_227_1780_929.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing Accuracy vs Acquired dataset size for Tiny-ImageNet\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right-center\",\n            \"image_style\": \"Standard chart with grid lines, legend, and axis labels\",\n            \"container\": null,\n            \"special_instructions\": \"The chart displays multiple line series with different markers and line styles. The legend is located in the bottom-right of the chart area. The x-axis is labeled 'Acquired dataset size' and the y-axis 'Accuracy'. The data points show a steady increase in accuracy.\"\n          }\n        },\n        {\n          \"id\": \"body-41-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"✓ Tiny-ImageNet\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2B2B2B\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right, below the second chart\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Text is preceded by a checkmark icon. Positioned directly below the right chart.\"\n          }\n        },\n        {\n          \"id\": \"body-41-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SAMSUNG SDS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2B2B2B\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right corner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold text, likely a footer or branding element.\"\n          }\n        },\n        {\n          \"id\": \"body-41-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"41\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2B2B2B\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right corner, below 'SAMSUNG SDS'\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small page number.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/58cd9988-f0f7-46c2-8e88-50af0ef33ddc_input_0.jpg", "images/multi_image/58cd9988-f0f7-46c2-8e88-50af0ef33ddc_input_1.jpg"], "target_image": "images/multi_image/58cd9988-f0f7-46c2-8e88-50af0ef33ddc_target.png", "width": 1920, "height": 1080}
+{"index": 123, "id": "6b7810c1-76fd-4274-a19c-5f8dc492a98f", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Times New Roman\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#6F5C93\",\n      \"secondary\": \"#006A62\",\n      \"text\": \"#000000\",\n      \"accent\": \"#B92B27\"\n    },\n    \"special_instructions\": \"The slide features a clean, white background typical of academic presentations. The design relies on color coding to distinguish between different mathematical concepts: purple for the main title, teal for the subtitle, and black for standard text. The diagram uses specific colors (green, black, blue, red) to map variables to visual elements.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Asymmetric, Two-column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 2,\n      \"blocks\": [\n        {\n          \"id\": \"title-2\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Background\",\n          \"style\": {\n            \"color\": \"#6F5C93\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold sans-serif font in purple.\"\n          }\n        },\n        {\n          \"id\": \"body-2-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Neural ordinary differential equations for time series modeling\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#006A62\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Teal colored sans-serif text, slightly smaller than title.\"\n          }\n        },\n        {\n          \"id\": \"equation-2-1\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"h(t_i) = c(x_i, h(t_{i-1}))\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#000000\",\n            \"position\": \"left column, below subtitle\",\n            \"container\": null,\n            \"special_instructions\": \"Standard serif math font.\"\n          }\n        },\n        {\n          \"id\": \"equation-2-2\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"x(t) = g(h(t))\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#000000\",\n            \"position\": \"left column, below previous equation\",\n            \"container\": null,\n            \"special_instructions\": \"Standard serif math font.\"\n          }\n        },\n        {\n          \"id\": \"equation-2-3\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"h(t) = h(t_i) + \\\\int_{t_i}^{t} f(h(s), s) ds\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#000000\",\n            \"position\": \"left column, bottom of text block\",\n            \"container\": null,\n            \"special_instructions\": \"Standard serif math font.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1160_690_2740_1469.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1160_690_2740_1469.jpg\"\n              }\n            ],\n            \"source\": \"Diagram illustrating time series modeling with ODEs\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, vertically centered\",\n            \"image_style\": \"Line art diagram with distinct colors for different curves and markers.\",\n            \"container\": null,\n            \"special_instructions\": \"The diagram shows a timeline at the bottom. A black curve labeled x(t) represents the observed data. A green curve labeled h(t) represents the hidden state. Blue diamonds mark specific time steps (t0, t1, t2). A red circle highlights the 'Update' process at t0, showing the transition from h(t) to the next state. Arrows indicate the flow of updates and the mapping function g(.).\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/6b7810c1-76fd-4274-a19c-5f8dc492a98f_input_0.jpg"], "target_image": "images/multi_image/6b7810c1-76fd-4274-a19c-5f8dc492a98f_target.png", "width": 2880, "height": 1620}
+{"index": 124, "id": "39bb5521-bee9-492d-8d37-30c3748b7d76", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#5B9BD5\",\n      \"secondary\": \"#2F5597\",\n      \"text\": \"#333333\",\n      \"accent\": \"#5B9BD5\"\n    },\n    \"special_instructions\": \"Clean, high-contrast academic slide design. Features a solid white background with a prominent blue footer banner. The main content area is uncluttered, focusing on two large data visualization plots separated by a conceptual arrow graphic. The visual style is strictly functional and technical, typical of computer science conferences like ICLR.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Technical, Professional, Minimalist\",\n    \"composition_style\": \"Asymmetric Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 4,\n      \"blocks\": [\n        {\n          \"id\": \"title-4\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"3-Armed Spiral\",\n          \"style\": {\n            \"color\": \"#333333\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold sans-serif font. Positioned in the upper-left quadrant of the slide.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large, hollow, chevron-shaped arrow pointing to the right. The arrow is outlined in a medium blue color and has no fill.\",\n          \"style\": {\n            \"illustration_style\": \"line art\",\n            \"color_scheme\": \"medium blue outline (#5B9BD5)\",\n            \"position\": \"center of the slide, vertically aligned between the two plots\",\n            \"special_instructions\": \"Acts as a visual connector indicating a transformation or process flow from the left plot to the right plot.\"\n          }\n        },\n        {\n          \"id\": \"body-4-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Biased data collect\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#5B9BD5\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left of the central arrow illustration\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small sans-serif text in blue, labeling the process represented by the arrow.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_41_286_999_1227.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_41_286_999_1227.jpg\"\n              }\n            ],\n            \"source\": \"3-Armed Spiral dataset visualization showing three distinct spiral arms in green, blue, and purple.\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left half of the slide\",\n            \"image_style\": \"Standard matplotlib plot style with grid lines and axis labels.\",\n            \"container\": null,\n            \"special_instructions\": \"The plot displays a 3D spiral projected onto 2D space with three arms of different colors.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1428_288_2381_1228.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1428_288_2381_1228.jpg\"\n              }\n            ],\n            \"source\": \"Resulting visualization after biased data collection, showing only two spiral arms (purple and blue).\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right half of the slide\",\n            \"image_style\": \"Standard matplotlib plot style with grid lines and axis labels.\",\n            \"container\": null,\n            \"special_instructions\": \"The plot displays a 2D spiral with two arms, visually demonstrating the effect of the 'Biased data collect' process.\"\n          }\n        },\n        {\n          \"id\": \"footer-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"UNC Department of Computer Science | Practical Integration via Separable Bijective Networks | ICLR 2022\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom of the slide, inside the blue banner\",\n            \"container\": \"solid blue banner (#5B9BD5) spanning full width\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"White text on a blue background. Left side contains the university logo and department name. Center contains the paper title. Right side contains the conference name.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/39bb5521-bee9-492d-8d37-30c3748b7d76_input_0.jpg", "images/multi_image/39bb5521-bee9-492d-8d37-30c3748b7d76_input_1.jpg"], "target_image": "images/multi_image/39bb5521-bee9-492d-8d37-30c3748b7d76_target.png", "width": 2427, "height": 1365}
+{"index": 125, "id": "c149bce8-1977-474e-9c17-de0681e6aff1", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Research Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#4A90E2\",\n      \"secondary\": \"#F5D0B0\",\n      \"text\": \"#000000\",\n      \"accent\": \"#78E08F\"\n    },\n    \"special_instructions\": \"Clean, high-contrast academic style. The slide relies on a stark white background with black text for maximum readability. Visual elements use a specific palette: a medium-light blue for charts, a soft peach/beige for diagram containers, and a mint green for highlights. The layout is dense, utilizing a two-column structure to fit complex data (tables and diagrams) alongside explanatory text.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Technical, Professional, Information-Dense\",\n    \"composition_style\": \"Asymmetric Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 14,\n      \"blocks\": [\n        {\n          \"id\": \"title-14\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Experiments\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold font size, positioned at the very top left corner.\"\n          }\n        },\n        {\n          \"id\": \"body-14-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"■ Training dataset\\nDataset for knowledge encoding: PathKT;\\nDatasets for KEP pretraining: OpenPath and Quilt1M\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style. The first bullet is a solid square. Key terms 'PathKT', 'OpenPath', and 'Quilt1M' are bolded.\"\n          }\n        },\n        {\n          \"id\": \"body-14-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"☑ Downstream tasks and datasets\\n• Retrieval (cross-modal retrieval, disease retrieval)\\n• Zero-shot patch classification\\n• Zero-shot WSI tumor subtyping\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, below first body block\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style. The first bullet is a solid square with a checkmark inside. Sub-points use standard circular bullets.\"\n          }\n        },\n        {\n          \"id\": \"table-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_table_170_880_1545_1515.jpg\",\n                \"relative_path\": \"imgs/img_in_table_170_880_1545_1515.jpg\"\n              }\n            ],\n            \"source\": \"Table detailing datasets for Retrieval, Classification, and Subtyping tasks\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"middle-left\",\n            \"image_style\": \"Standard table with thin black borders and horizontal dividers.\",\n            \"container\": null,\n            \"special_instructions\": \"The table is aligned to the left column, spanning vertically from the middle to the bottom of the slide.\"\n          }\n        },\n        {\n          \"id\": \"chart-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1641_228_2856_498.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1641_228_2856_498.jpg\"\n              }\n            ],\n            \"source\": \"Bar chart showing disease distribution\",\n            \"caption\": \"The disease distribution in PathKT\"\n          },\n          \"style\": {\n            \"position\": \"top-right\",\n            \"image_style\": \"Vertical bar chart with light blue bars (#87CEEB) and black text labels.\",\n            \"container\": null,\n            \"special_instructions\": \"Located in the upper right quadrant. The caption is centered below the chart.\"\n          }\n        },\n        {\n          \"id\": \"caption-14-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"The disease distribution in PathKT\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"below chart-1\",\n            \"text_style\": null,\n            \"special_instructions\": null\n          }\n        },\n        {\n          \"id\": \"illustration-14-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A schematic diagram illustrating a 'Zero-shot WSI tumor subtyping' pipeline. It features a large histology image on the left with a grid overlay, arrows pointing to smaller patch images, which feed into a 'Zero-shot classifier'. The output flows through 'Top-K pooling' to a final 'Final diagnosis' label. The diagram uses peach-colored rounded rectangles for containers and mint-green squares for probability values.\",\n          \"style\": {\n            \"illustration_style\": \"Conceptual flowchart/diagram\",\n            \"color_scheme\": \"Peach (#F5D0B0) for containers, Mint Green (#78E08F) for highlights, Black for text and arrows\",\n            \"position\": \"bottom-right\",\n            \"special_instructions\": \"The diagram is positioned in the bottom right corner, below the bar chart.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/c149bce8-1977-474e-9c17-de0681e6aff1_input_0.jpg", "images/multi_image/c149bce8-1977-474e-9c17-de0681e6aff1_input_1.jpg", "images/multi_image/c149bce8-1977-474e-9c17-de0681e6aff1_input_2.jpg"], "target_image": "images/multi_image/c149bce8-1977-474e-9c17-de0681e6aff1_target.png", "width": 2880, "height": 1620}
+{"index": 126, "id": "00cae2e8-1841-44ce-b5f9-da55832e9d38", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#D3DCE6\",\n      \"secondary\": \"#E6E6E6\",\n      \"text\": \"#000000\",\n      \"accent\": \"#000000\"\n    },\n    \"special_instructions\": \"The slide features a clean, minimalist academic aesthetic with a stark white background. The design relies heavily on typography and mathematical notation. Visual elements are rendered in flat, solid colors (grey for diagrams, green for arrows). There are no gradients or complex textures. The layout is structured into two distinct columns, separated by whitespace.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Two-column Asymmetric\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 4,\n      \"blocks\": [\n        {\n          \"id\": \"title-4\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"1-min Highlight: Methodology\\nSeparability (SupCon) and Alignment (IRM)\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Two-line title. The first line is slightly larger than the second. No background container.\"\n          }\n        },\n        {\n          \"id\": \"body-4-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Challenge 1: Separability\\nSupervised Contrastive Learning (SupCon)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-column, top\",\n            \"container\": {\n              \"color\": \"#D3DCE6\",\n              \"shape\": \"rounded rectangle\",\n              \"width\": \"full column width\"\n            },\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is placed directly below a light blue-grey background banner.\"\n          }\n        },\n        {\n          \"id\": \"equation-4-1\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"$$ \\\\phi: \\\\mathcal{X} \\\\mapsto \\\\mathbb{R}^{m} $$\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#000000\",\n            \"position\": \"left-column, middle\",\n            \"container\": null,\n            \"special_instructions\": \"Standard LaTeX math font.\"\n          }\n        },\n        {\n          \"id\": \"equation-4-2\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"$$ \\\\mathcal{L}_{SupCon} = -\\\\frac{1}{|\\\\mathcal{P}(\\\\mathbf{x})|}\\\\sum_{\\\\mathbf{z}^{+}\\\\in\\\\mathcal{P}(\\\\mathbf{x})}\\\\log\\\\frac{\\\\exp\\\\left(\\\\mathbf{z}^{\\\\top}\\\\cdot\\\\mathbf{z}^{+}/\\\\tau\\\\right)}{\\\\sum_{\\\\mathbf{z}^{-}\\\\in\\\\mathcal{N}(\\\\mathbf{x})}\\\\exp\\\\left(\\\\mathbf{z}^{\\\\top}\\\\cdot\\\\mathbf{z}^{-}/\\\\tau\\\\right)} $$\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#000000\",\n            \"position\": \"left-column, bottom\",\n            \"container\": null,\n            \"special_instructions\": \"Standard LaTeX math font.\"\n          }\n        },\n        {\n          \"id\": \"image-4-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_298_842_1715_1550.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_298_842_1715_1550.jpg\"\n              }\n            ],\n            \"source\": \"Diagram illustrating the SupCon methodology\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left-column, vertically centered\",\n            \"image_style\": \"Flat vector graphics style\",\n            \"container\": null,\n            \"special_instructions\": \"Diagram shows a cat image (X) splitting into two augmentations, feeding into a CNN backbone (grey cube) and a penultimate layer (grey plane), resulting in a normalized vector Z (green arrow).\"\n          }\n        },\n        {\n          \"id\": \"body-4-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Challenge 2: Alignment\\nInvariant Risk Minimization (IRM)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"right-column, top\",\n            \"container\": {\n              \"color\": \"#D3DCE6\",\n              \"shape\": \"rounded rectangle\",\n              \"width\": \"full column width\"\n            },\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is placed directly below a light blue-grey background banner.\"\n          }\n        },\n        {\n          \"id\": \"equation-4-3\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"$$ \\\\operatorname*{min}_{\\\\phi,\\\\beta^{*}}\\\\frac{1}{|\\\\mathcal{E}|}\\\\sum_{e\\\\in\\\\mathcal{E}}\\\\mathcal{R}^{e}(\\\\phi,\\\\beta^{*})\\\\rightarrow\\\\mathcal{L}_{IRM} $$\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#000000\",\n            \"position\": \"right-column, upper-middle\",\n            \"container\": null,\n            \"special_instructions\": \"Standard LaTeX math font.\"\n          }\n        },\n        {\n          \"id\": \"equation-4-4\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"$$ \\\\text{s.t. } \\\\beta^{*} \\\\in \\\\arg \\\\min_{\\\\beta} \\\\mathcal{R}^{e}(\\\\phi, \\\\beta) \\\\quad \\\\forall e \\\\in \\\\mathcal{E} $$\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#000000\",\n            \"position\": \"right-column, middle\",\n            \"container\": {\n              \"color\": \"#E6E6E6\",\n              \"shape\": \"rounded rectangle\",\n              \"width\": \"full column width\"\n            },\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Equation is enclosed in a light grey box.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Two conceptual diagrams illustrating 'Violate' and 'Satisfied' states. Each diagram shows three columns of data points (crosses and circles) separated by dashed lines. The 'Violate' diagram has a red arrow pointing to it. The 'Satisfied' diagram has a green arrow pointing to it.\",\n          \"style\": {\n            \"illustration_style\": \"Conceptual diagram / schematic\",\n            \"color_scheme\": \"Blue, Green, Yellow for data points; Red and Green for arrows\",\n            \"position\": \"right-column, bottom\",\n            \"special_instructions\": \"The 'Violate' diagram shows misaligned clusters. The 'Satisfied' diagram shows aligned clusters. The arrows originate from the equation box above.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/00cae2e8-1841-44ce-b5f9-da55832e9d38_input_0.jpg", "images/multi_image/00cae2e8-1841-44ce-b5f9-da55832e9d38_input_1.jpg"], "target_image": "images/multi_image/00cae2e8-1841-44ce-b5f9-da55832e9d38_target.png", "width": 3840, "height": 2160}
+{"index": 127, "id": "84e7351e-92a6-4f91-a983-11fd5048dd98", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#0056B3\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#0056B3\"\n    },\n    \"special_instructions\": \"Clean, high-contrast academic style. The slide features a white background with black text. A distinct blue graphical element anchors the top-left corner. The chart uses a standard matplotlib default color cycle (Red, Blue, Green) with solid and dashed line styles to differentiate data series.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Informative\",\n    \"composition_style\": \"Asymmetric, Left-aligned text, Centered chart\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 14,\n      \"blocks\": [\n        {\n          \"id\": \"title-14\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"How the performance is improved?\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"The title is positioned to the right of a blue chevron shape. The text is large and bold.\"\n          }\n        },\n        {\n          \"id\": \"illustration-14-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A solid blue chevron (arrowhead pointing right) shape located in the top-left corner, partially overlapping the slide margin.\",\n          \"style\": {\n            \"illustration_style\": \"geometric shape\",\n            \"color_scheme\": \"Solid blue (#0056B3)\",\n            \"position\": \"top-left corner\",\n            \"special_instructions\": \"Acts as a design anchor for the slide number.\"\n          }\n        },\n        {\n          \"id\": \"slide-number-14\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"4\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"minor\",\n            \"position\": \"top-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"White text inside the blue chevron shape.\"\n          }\n        },\n        {\n          \"id\": \"separator-14\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A thin horizontal line extending from the right edge of the blue chevron across the top of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"line\",\n            \"color_scheme\": \"Blue (#0056B3)\",\n            \"position\": \"top, below title\",\n            \"special_instructions\": \"Matches the color of the chevron.\"\n          }\n        },\n        {\n          \"id\": \"body-14-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Assumption: a global modeling of all instances from one category\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-aligned, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text, standard weight.\"\n          }\n        },\n        {\n          \"id\": \"body-14-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Evidence: more useful on the cases with multiple instances from one category\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-aligned, below previous body text\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text, standard weight.\"\n          }\n        },\n        {\n          \"id\": \"figure-14-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_314_504_1629_1076.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_314_504_1629_1076.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing relative mAP change across dataset partitions\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-bottom\",\n            \"image_style\": \"standard plot frame with black axes\",\n            \"container\": null,\n            \"special_instructions\": \"The chart includes a legend with 6 entries distinguishing between solid and dashed lines in red, blue, and green.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/84e7351e-92a6-4f91-a983-11fd5048dd98_input_0.jpg"], "target_image": "images/multi_image/84e7351e-92a6-4f91-a983-11fd5048dd98_target.png", "width": 1920, "height": 1080}
+{"index": 128, "id": "b9486a08-f041-4b43-b063-2a0d1c43e1b2", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#000000\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#88CC88\"\n    },\n    \"special_instructions\": \"The slide features a stark, high-contrast academic aesthetic. The background is pure white. The design relies on standard sans-serif typography. The only color variation comes from the diagram on the right, which uses a soft green palette for the neural network nodes and light blue for the text tokens, contrasting with the black text on the left.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Technical, Professional, Minimalist\",\n    \"composition_style\": \"Asymmetric, Two-column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 41,\n      \"blocks\": [\n        {\n          \"id\": \"title-41\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Implicit Language Q Learning – two solutions\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold title aligned to the top-left corner.\"\n          }\n        },\n        {\n          \"id\": \"body-41-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"We can fix this by either pushing down OOD probabilities or Q values.\\n\\no Probabilities: add top-p filter or temperature to the logits\\n\\no Q values: add NLL loss to the Q-values.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, top\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points with circular markers. Indented sub-points.\"\n          }\n        },\n        {\n          \"id\": \"body-41-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Both work in practice. We find the latter typically requires the least amount of tuning.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard bullet point.\"\n          }\n        },\n        {\n          \"id\": \"body-41-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"ILQL = IQL loss + CQL loss\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Equation-style text, no bullet point.\"\n          }\n        },\n        {\n          \"id\": \"illustration-41-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A complex architectural diagram illustrating an Implicit Language Q Learning model. It features a 'Task Data' cylinder on the left feeding into a 'SL' block. This connects to a 'pi_theta Transformer' at the bottom and a 'Transformer Value Function' in the middle. The diagram includes green nodes representing tokens (e.g., 'Hi', 'How', 'are') with associated Q and V values, and blue tokens representing text sequences. Arrows indicate data flow, with a 'Copy' operation leading to a 'Target Net'. A zoomed-in box on the top left details the 'IQL: Implicit Backup' loss functions.\",\n          \"style\": {\n            \"illustration_style\": \"flowchart\",\n            \"color_scheme\": \"Green nodes, blue tokens, grey boxes, black lines\",\n            \"position\": \"right side, spanning vertically\",\n            \"special_instructions\": \"The diagram uses a clean, flat design style typical of technical papers.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/b9486a08-f041-4b43-b063-2a0d1c43e1b2_input_0.jpg"], "target_image": "images/multi_image/b9486a08-f041-4b43-b063-2a0d1c43e1b2_target.png", "width": 1440, "height": 810}
+{"index": 129, "id": "ed41c5e9-0c8d-496d-8f40-2a8d09d2d5d4", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Research Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#F25022\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#F25022\"\n    },\n    \"special_instructions\": \"The slide features a clean, professional academic aesthetic. The top of the slide is dominated by a solid, vibrant orange banner (#F25022) that spans the full width, containing the main title in white. The rest of the slide is a stark white background with black text. The layout is asymmetric, with a bulleted list on the left and a data table on the right. The typography is standard sans-serif (likely Arial), ensuring high readability for technical content.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Academic\",\n    \"composition_style\": \"Asymmetric, Two-column layout (Text Left, Table Right)\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 13,\n      \"blocks\": [\n        {\n          \"id\": \"title-13\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Experiments\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left, inside the orange banner\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Located within a full-width solid orange (#F25022) banner at the very top of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-13-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Ablation Studies\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left, below the orange banner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Section header, bolded, larger font size than body text.\"\n          }\n        },\n        {\n          \"id\": \"body-13-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"➤ Fusion Layer\\n\\n➤ Enhance the memorization\\n\\n➤ Weaken the generalization\\n\\n➤ Performance is even degrade\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bulleted list using a right-pointing triangle bullet. Standard font weight.\"\n          }\n        },\n        {\n          \"id\": \"body-13-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"➤ Fusion Layer + Feature-wise Mask\\n\\n➤ Solve the generalization to a certain degree\\n\\n➤ Performance is slightly improved\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bulleted list using a right-pointing triangle bullet.\"\n          }\n        },\n        {\n          \"id\": \"body-13-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"➤ Fusion Layer + Feature-wise Mask + Adaptive\\n\\n➤ Well balance of memorization and generalization\\n\\n➤ Performance is significantly improved\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bulleted list using a right-pointing triangle bullet.\"\n          }\n        },\n        {\n          \"id\": \"caption-13-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Table 2: Ablation study. Based on the baseline (BL), Fusion (Fu), Feature-wise Mask (FwM), and Adaptive (Ada) is gradually added for ablation studies. ↑ means higher is better and ↓ means lower is better.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"top-left of the table area\",\n            \"text_style\": null,\n            \"special_instructions\": \"Small text caption above the table, left-aligned relative to the table.\"\n          }\n        },\n        {\n          \"id\": \"table-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_table_1290_692_2393_1160.jpg\",\n                \"relative_path\": \"imgs/img_in_table_1290_692_2393_1160.jpg\"\n              }\n            ],\n            \"source\": \"Ablation study results table\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"middle-right\",\n            \"image_style\": \"Standard table with grid lines, bold headers, and bolded best values.\",\n            \"container\": null,\n            \"special_instructions\": \"A detailed data table comparing model performance across different configurations (BL, Fu, FwM, Ada) and datasets (Criteo, Terabyte, Avazu). The table uses checkmarks to indicate included features and bold text to highlight the best results.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/ed41c5e9-0c8d-496d-8f40-2a8d09d2d5d4_input_0.jpg"], "target_image": "images/multi_image/ed41c5e9-0c8d-496d-8f40-2a8d09d2d5d4_target.png", "width": 2592, "height": 1458}
+{"index": 130, "id": "a157c6e1-d331-4766-9680-2158e3c63760", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#C41E3A\",\n      \"secondary\": \"#555555\",\n      \"text\": \"#555555\",\n      \"accent\": \"#C41E3A\"\n    },\n    \"special_instructions\": \"Clean, high-contrast academic style. White background with dark grey body text and a distinct red accent color for the main title. The layout is split horizontally, with text on the left and a complex technical diagram on the right. The diagram uses a mix of greyscale and specific colors (green, red, blue) to denote different data flows and regions.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Informative\",\n    \"composition_style\": \"Asymmetric, Split Layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 18,\n      \"blocks\": [\n        {\n          \"id\": \"title-18\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Key properties\",\n          \"style\": {\n            \"color\": \"#C41E3A\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold red text located in the upper left corner.\"\n          }\n        },\n        {\n          \"id\": \"body-18-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"The nodes in our method are\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#555555\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Introductory sentence in dark grey.\"\n          }\n        },\n        {\n          \"id\": \"body-18-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"1. dynamic - adapt according to the saliency of the input at each time step.\\n2. localized - pool from clearly delimited area in space + maintain position information\\n3. differentiable - w.r.t. regions' params, allowing us to learn regions from the final loss, w/o direct supervision for the regions.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#555555\",\n            \"importance\": \"supporting\",\n            \"position\": \"below intro sentence, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Numbered list. The keywords 'dynamic', 'localized', and 'differentiable' are bolded.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_545_714_2347_1510.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_545_714_2347_1510.jpg\"\n              }\n            ],\n            \"source\": \"Technical diagram illustrating the method's flow from input image X_t to output Y_t\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-right, spanning the lower half of the slide\",\n            \"image_style\": \"No border or shadow\",\n            \"container\": null,\n            \"special_instructions\": \"Complex diagram showing a basketball player image being processed through several stages involving feature extraction (o1, v1), graph connections, and region pooling (k1) to produce a saliency map (Y_t). The diagram includes mathematical notation and colored dots representing specific regions.\"\n          }\n        },\n        {\n          \"id\": \"caption-18-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Discovering Dynamic Salient Regions for Spatio-Temporal Graph Neural Networks - Neurips 2021\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#555555\",\n            \"position\": \"bottom-left\",\n            \"text_style\": null,\n            \"special_instructions\": \"Small footer text indicating the paper title and conference.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/a157c6e1-d331-4766-9680-2158e3c63760_input_0.jpg"], "target_image": "images/multi_image/a157c6e1-d331-4766-9680-2158e3c63760_target.png", "width": 2880, "height": 1620}
+{"index": 131, "id": "7849920a-7809-402e-9c9e-c2373083d40e", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2E4C91\",\n      \"secondary\": \"#8B0000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#2E4C91\"\n    },\n    \"special_instructions\": \"The slide features a clean, high-contrast academic style. The background is stark white. A solid, deep blue horizontal banner spans the top edge, containing the title. The main content area is dominated by a line chart with a light grey grid. The chart uses specific colors for data series: a bright blue for the rising curve, a dark red for the decaying curve, and black for the dashed bound line. Text is black, with the figure caption in the same deep blue as the header banner.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Technical, Professional, Data-Driven\",\n    \"composition_style\": \"Top-aligned header, Centered chart, Bottom-aligned caption\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 31,\n      \"blocks\": [\n        {\n          \"id\": \"title-31\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Invariance gap at different optima\",\n          \"style\": {\n            \"color\": \"#2E4C91\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left, inside a blue banner\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Text is white on a solid deep blue (#2E4C91) background bar that spans the full width of the slide.\"\n          }\n        },\n        {\n          \"id\": \"chart-31\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_459_295_1947_1099.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_459_295_1947_1099.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing invariance gap vs dimensions\",\n            \"caption\": \"Figure: N = 10 dependent data points.\"\n          },\n          \"style\": {\n            \"position\": \"centered horizontally and vertically in the main content area\",\n            \"image_style\": \"Standard matplotlib style with light grey grid lines, black axes, and a legend box in the upper right quadrant of the plot area.\",\n            \"container\": null,\n            \"special_instructions\": \"The chart displays three lines: a blue line rising sharply, a red line peaking and decaying, and a black dashed horizontal line. The legend is located inside the plot area.\"\n          }\n        },\n        {\n          \"id\": \"caption-31-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Figure: N = 10 dependent data points.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2E4C91\",\n            \"position\": \"bottom-center, below the chart\",\n            \"text_style\": null,\n            \"special_instructions\": \"The text color matches the header banner color.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/7849920a-7809-402e-9c9e-c2373083d40e_input_0.jpg"], "target_image": "images/multi_image/7849920a-7809-402e-9c9e-c2373083d40e_target.png", "width": 2409, "height": 1365}
+{"index": 132, "id": "a5c27f19-508f-4ed2-a026-248a1f072201", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Tech Network Dark\",\n    \"background\": {\n      \"color\": \"#0A2E45\",\n      \"gradient\": \"linear-gradient(135deg, #0A2E45 0%, #051622 100%)\",\n      \"texture\": \"none\",\n      \"pattern\": \"subtle geometric network lines and glowing orbs\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#FFFFFF\",\n      \"secondary\": \"#E6E6E6\",\n      \"text\": \"#FFFFFF\",\n      \"accent\": \"#D4D400\"\n    },\n    \"special_instructions\": \"The slide features a dark, deep blue background with a subtle, low-opacity geometric network pattern (lines and nodes) overlaid. There are soft, glowing orbs of light (cyan and red) in the corners, adding depth. The overall feel is technical and modern.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Technical, Professional\",\n    \"composition_style\": \"Asymmetric, Two-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 6,\n      \"blocks\": [\n        {\n          \"id\": \"title-6\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"BASiS Algorithm\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold white text at the top center of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-6-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Initialization:\\n• Define an anchor set\\n• Define the reference embedding space\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points with standard white text. The main header 'Initialization:' is bolded.\"\n          }\n        },\n        {\n          \"id\": \"body-6-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Training loop:\\n• Calculate the embedding space over a new batch\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#D4D400\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points with yellow/gold text. The main header 'Training loop:' is bolded.\"\n          }\n        },\n        {\n          \"id\": \"equation-6-1\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"V^b \\\\rightarrow \\\\varphi \\\\quad V^a \\\\rightarrow \\\\varphi^a\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"position\": \"top-right\",\n            \"container\": null,\n            \"special_instructions\": \"Mathematical notation displayed above the top chart.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_870_252_1312_436.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_870_252_1312_436.jpg\"\n              }\n            ],\n            \"source\": \"Scatter plot showing three curved clusters of points (red, blue, green) on a grid.\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right-middle\",\n            \"image_style\": \"standard\",\n            \"container\": null,\n            \"special_instructions\": \"A scatter plot image showing data points arranged in three curved shapes.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_899_465_1284_742.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_899_465_1284_742.jpg\"\n              }\n            ],\n            \"source\": \"Scatter plot showing three distinct points (red, blue, green) on a Cartesian coordinate system.\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right-bottom\",\n            \"image_style\": \"standard\",\n            \"container\": null,\n            \"special_instructions\": \"A scatter plot image showing three distinct colored squares on a grid with axes labeled u1 and u2.\"\n          }\n        },\n        {\n          \"id\": \"caption-6-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"6\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"position\": \"bottom-right\",\n            \"text_style\": null,\n            \"special_instructions\": \"Slide number in the bottom right corner.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/a5c27f19-508f-4ed2-a026-248a1f072201_input_0.jpg", "images/multi_image/a5c27f19-508f-4ed2-a026-248a1f072201_input_1.jpg"], "target_image": "images/multi_image/a5c27f19-508f-4ed2-a026-248a1f072201_target.png", "width": 1440, "height": 810}
+{"index": 133, "id": "cbcd44e4-1d7e-4684-a67d-727c59b0e6b8", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#4F66B5\",\n      \"secondary\": \"#99AACC\",\n      \"text\": \"#333333\",\n      \"accent\": \"#4F66B5\"\n    },\n    \"special_instructions\": \"The slide features a clean, minimalist academic design. The primary visual element is a header banner at the top and bottom. The top banner is a solid block of color (#99AACC) containing a darker title text. The bottom banner is a solid block of color (#4F66B5) containing white text. The main content area is stark white. Typography is strictly sans-serif (Arial), utilizing bolding for emphasis and italics for mathematical notation. The layout is centered and symmetrical.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Minimalist\",\n    \"composition_style\": \"Centered, Symmetrical\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 19,\n      \"blocks\": [\n        {\n          \"id\": \"title-19\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Barycenters of continuous measures\",\n          \"style\": {\n            \"color\": \"#4F66B5\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left, below the top banner\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, dark blue text. Serves as the main heading for the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-19-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Barycenter of 5 Gaussian distributions with mean and covariance randomly generated.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center, below the title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard body text, left-aligned relative to the slide width, describing the experimental setup.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_133_170_359_333.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_133_170_359_333.jpg\"\n              },\n              {\n                \"path\": \"imgs/img_in_image_box_365_171_591_330.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_365_171_591_330.jpg\"\n              }\n            ],\n            \"source\": \"Scatter plot comparison showing method output vs true Wasserstein barycenter\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center, horizontally spanning the middle of the slide\",\n            \"image_style\": \"Two distinct scatter plots displayed side-by-side. Both feature blue dots on a light blue gradient background. The left plot shows a circular distribution, the right shows an elliptical distribution.\",\n            \"container\": null,\n            \"special_instructions\": \"The images are embedded directly into the slide layout without visible borders or captions underneath them.\"\n          }\n        },\n        {\n          \"id\": \"body-19-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"scatter plot: output of our method\\nlevel sets of its density: true Wasserstein barycenter\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left, below the images\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Two lines of text. The first line starts with 'scatter plot:' in bold blue (#4F66B5). The second line starts with 'level sets of its density:' in bold blue (#4F66B5). The rest of the text is black.\"\n          }\n        },\n        {\n          \"id\": \"equation-19-1\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"FW recovers both the mean and covariance of the target barycenter.\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#4F66B5\",\n            \"position\": \"bottom-left, below the explanatory text\",\n            \"container\": null,\n            \"special_instructions\": \"Italicized text in the primary accent color (#4F66B5), functioning as a concluding statement or theorem.\"\n          }\n        },\n        {\n          \"id\": \"footer-19\",\n          \"type\": \"text\",\n          \"role\": \"footer\",\n          \"content\": \"Sinkhorn Barycenters via Frank Wolfe | 19 / 22\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right, inside the bottom banner\",\n            \"container\": \"solid banner\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"White text on a solid blue background. Left part is the slide title, right part is the slide number.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/cbcd44e4-1d7e-4684-a67d-727c59b0e6b8_input_0.jpg", "images/multi_image/cbcd44e4-1d7e-4684-a67d-727c59b0e6b8_input_1.jpg"], "target_image": "images/multi_image/cbcd44e4-1d7e-4684-a67d-727c59b0e6b8_target.png", "width": 726, "height": 545}
+{"index": 134, "id": "f780eddd-0e5f-459d-ba82-3711d77892b7", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Research Results\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2B4C7E\",\n      \"secondary\": \"#666666\",\n      \"text\": \"#333333\",\n      \"accent\": \"#F4C4A2\"\n    },\n    \"special_instructions\": \"The slide features a clean, minimalist academic design. The background is stark white. Typography is strictly sans-serif (likely Arial). A thin, dark blue horizontal line separates the title from the body content. The charts utilize a soft, pastel color palette (peach, light blue, grey, light green) to differentiate data points without harsh contrast.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Data-Driven\",\n    \"composition_style\": \"Top-down hierarchy with a three-column grid layout for the charts at the bottom.\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 21,\n      \"blocks\": [\n        {\n          \"id\": \"title-21\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Empirical Results\",\n          \"style\": {\n            \"color\": \"#2B4C7E\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold text aligned to the left margin. Followed immediately by a thin horizontal rule.\"\n          }\n        },\n        {\n          \"id\": \"body-21-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"How to classify edges using the TDLG?\\nSimply use rows of A as temporal edge embeddings.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard body text. The variable 'A' is bolded.\"\n          }\n        },\n        {\n          \"id\": \"body-21-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"This achieves good classification performance across several real-world datasets:\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#CCCCCC\",\n            \"importance\": \"minor\",\n            \"position\": \"centered horizontally, above the charts\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Text is rendered in a very light grey, almost blending into the white background, serving as a subtle lead-in to the visual data.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_153_845_1271_1356.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_153_845_1271_1356.jpg\"\n              }\n            ],\n            \"source\": \"Bar chart comparing Mean Test AUC (%) for Temporal Edge Classification across multiple models.\",\n            \"caption\": \"Temporal Edge Classification\"\n          },\n          \"style\": {\n            \"position\": \"bottom-left\",\n            \"image_style\": \"Standard bar chart with grid lines.\",\n            \"container\": null,\n            \"special_instructions\": \"The chart displays TDLG as the top performer with a peach-colored bar.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1288_846_2000_1354.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1288_846_2000_1354.jpg\"\n              }\n            ],\n            \"source\": \"Bar chart comparing performance for Interpolative Link Prediction.\",\n            \"caption\": \"Interpolative Link Prediction\"\n          },\n          \"style\": {\n            \"position\": \"bottom-center\",\n            \"image_style\": \"Standard bar chart with grid lines.\",\n            \"container\": null,\n            \"special_instructions\": \"The chart shows TDLG significantly outperforming other models like CTDNE and TIMERS.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_2016_845_2722_1356.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_2016_845_2722_1356.jpg\"\n              }\n            ],\n            \"source\": \"Bar chart comparing performance for Extrapolative Link Prediction.\",\n            \"caption\": \"Extrapolative Link Prediction\"\n          },\n          \"style\": {\n            \"position\": \"bottom-right\",\n            \"image_style\": \"Standard bar chart with grid lines.\",\n            \"container\": null,\n            \"special_instructions\": \"Similar to the previous chart, TDLG (peach bar) shows the highest AUC score.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/f780eddd-0e5f-459d-ba82-3711d77892b7_input_0.jpg", "images/multi_image/f780eddd-0e5f-459d-ba82-3711d77892b7_input_1.jpg", "images/multi_image/f780eddd-0e5f-459d-ba82-3711d77892b7_input_2.jpg"], "target_image": "images/multi_image/f780eddd-0e5f-459d-ba82-3711d77892b7_target.png", "width": 2880, "height": 1620}
+{"index": 135, "id": "777a0448-bb5a-4af3-b161-0f77979dd07a", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#0055A4\",\n      \"secondary\": \"#D9534F\",\n      \"text\": \"#000000\",\n      \"accent\": \"#E6B800\"\n    },\n    \"special_instructions\": \"Clean, high-contrast academic slide design. Uses a white background with black text for readability. Key terms and loss function names are highlighted with a dark red background (#D9534F) and white text. The layout is split into two distinct columns. The 'General Pipeline' section uses a schematic diagram style with solid lines and icons. The 'Calibration Losses' section uses a bar chart visualization with orange bars and colored check/cross marks.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Informative\",\n    \"composition_style\": \"Two-column asymmetric layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 24,\n      \"blocks\": [\n        {\n          \"id\": \"title-24\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"(Visual) Sim2Real Transfer\\n\\nAUGCAL: \\\"Proactive\\\" Sim2Real Calibrated Adaptation\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold font. The first line is slightly larger than the second.\"\n          }\n        },\n        {\n          \"id\": \"body-24-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"General Pipeline\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#0055A4\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, top\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, blue color matching the primary accent.\"\n          }\n        },\n        {\n          \"id\": \"illustration-24-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A schematic diagram illustrating a machine learning pipeline. It shows two input sources: 'Sim (GTAV)' and 'Real (Cityscapes)'. The Sim input goes through an 'AUG' (Augmentation) step. Both augmented Sim data and Real data feed into a 'Model' (represented by a neural network icon with a shield). From the model, three outputs branch out: 'Source Loss' (blue arrow), 'Adaptation Loss' (red arrow), and 'CALibration Loss' (dashed black arrow).\",\n          \"style\": {\n            \"illustration_style\": \"schematic diagram\",\n            \"color_scheme\": \"Blue, Red, Black, White\",\n            \"position\": \"left column, vertically centered\",\n            \"special_instructions\": \"Uses simple geometric shapes for images and arrows. Text labels are in black with red headers for loss types.\"\n          }\n        },\n        {\n          \"id\": \"body-24-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Calibration Losses\\n\\nModulate your classification scores at a minibatch level to be more calibrated\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, top\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The heading 'Calibration Losses' is bold and larger. The description text is smaller and left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"equation-24-1\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"\\\\sum_{batch} | \\\\text{Confidence} - \\\\text{Correctness} |\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#000000\",\n            \"position\": \"right column, middle\",\n            \"container\": null,\n            \"special_instructions\": \"Large mathematical notation. The summation symbol is large and grey. The text 'Confidence - Correctness' is inside absolute value bars.\"\n          }\n        },\n        {\n          \"id\": \"image-24-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_2094_888_3765_1747.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_2094_888_3765_1747.jpg\"\n              }\n            ],\n            \"source\": \"Calibration Loss visualization chart\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, below equation\",\n            \"image_style\": \"Bar chart with orange bars\",\n            \"container\": null,\n            \"special_instructions\": \"A bar chart showing 'Confidence' vs 'Correctness' for 7 images. Orange bars represent confidence scores. Red 'X' and blue checkmarks indicate correctness.\"\n          }\n        },\n        {\n          \"id\": \"body-24-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"DCA (Liang et. al, BMVC 2020)\\nMDCA (Hebbalgoupe et. al, CVPR 2022)\\nMbLS (Liu et. al, CVPR 2022)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"right column, bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"List of related works, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"illustration-24-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large grey curly brace on the left side of the chart and a matching one on the right side, visually grouping the 7 image rows together.\",\n          \"style\": {\n            \"illustration_style\": \"decorative bracket\",\n            \"color_scheme\": \"Grey\",\n            \"position\": \"flanking the chart on the right column\",\n            \"special_instructions\": \"Thin, light grey lines.\"\n          }\n        },\n        {\n          \"id\": \"illustration-24-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large grey summation symbol (Sigma) with 'batch' subscript to the left of the curly brace.\",\n          \"style\": {\n            \"illustration_style\": \"mathematical symbol\",\n            \"color_scheme\": \"Grey\",\n            \"position\": \"far left of the right column, vertically centered with the chart\",\n            \"special_instructions\": \"Large font size.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/777a0448-bb5a-4af3-b161-0f77979dd07a_input_0.jpg", "images/multi_image/777a0448-bb5a-4af3-b161-0f77979dd07a_input_1.jpg"], "target_image": "images/multi_image/777a0448-bb5a-4af3-b161-0f77979dd07a_target.png", "width": 3840, "height": 2160}
+{"index": 136, "id": "82306d44-4614-42e9-bb7c-79663070dc9b", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#66BB6A\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#66BB6A\"\n    },\n    \"special_instructions\": \"Clean, minimalist academic style. High contrast black text on white background. Key terms are highlighted in a specific shade of green (#66BB6A). The layout is asymmetrical, utilizing the left side for a large illustrative image and the right side for bullet points.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Asymmetric, Two-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 6,\n      \"blocks\": [\n        {\n          \"id\": \"title-6\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Why do we need unsupervised domain adaptation for object detectors?\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"The word 'unsupervised' is highlighted in green (#66BB6A) to emphasize the specific topic.\"\n          }\n        },\n        {\n          \"id\": \"body-6-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Challenge 1: Object-level labels are expensive or even unavailable\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#66BB6A\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The phrase 'Challenge 1:' is highlighted in green (#66BB6A) to denote a specific point being discussed.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_212_357_1080_1006.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_212_357_1080_1006.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of object detection annotations\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"middle-left\",\n            \"image_style\": \"No border or shadow\",\n            \"container\": null,\n            \"special_instructions\": \"A photograph of a dog on grass with a teal segmentation mask overlay and a bounding box around smaller figures in the background.\"\n          }\n        },\n        {\n          \"id\": \"body-6-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Object-level annotation\\n\\n• Category of each object\\n  • Bounding box\\n  • Pixel-level mask\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard bullet point list. The main header 'Object-level annotation' is bolded. Indentation is used for sub-points.\"\n          }\n        },\n        {\n          \"id\": \"caption-6-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"6\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"bottom-right corner\",\n            \"text_style\": null,\n            \"special_instructions\": \"Slide number.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/82306d44-4614-42e9-bb7c-79663070dc9b_input_0.jpg"], "target_image": "images/multi_image/82306d44-4614-42e9-bb7c-79663070dc9b_target.png", "width": 1920, "height": 1080}
+{"index": 137, "id": "d178a4db-6542-48c9-961e-e63d7c4d6299", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Minimalist\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#00204D\",\n      \"secondary\": \"#5C5338\",\n      \"text\": \"#5C5338\",\n      \"accent\": \"#8B0000\"\n    },\n    \"special_instructions\": \"Clean, high-contrast academic presentation style. The slide relies on a stark white background to maximize the visibility of the technical diagram. The color palette is strictly limited to deep navy blue for the main title, a muted olive-brown for body text and diagram axes, and a deep red for key emphasis words. The diagram itself uses a flat, vector-style aesthetic with a light blue semi-transparent polygon.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Top-down hierarchy with centered alignment for text and a horizontally centered diagram below.\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 14,\n      \"blocks\": [\n        {\n          \"id\": \"title-14\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Given data without labels\",\n          \"style\": {\n            \"color\": \"#00204D\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, sans-serif font. Serves as the primary heading.\"\n          }\n        },\n        {\n          \"id\": \"body-14-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"What can we learn from unlabeled data? Halfspace!\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#5C5338\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Body text in muted olive-brown. The word 'Halfspace!' is emphasized in a darker red (#8B0000) to draw attention to the concept.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_300_336_1140_744.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_300_336_1140_744.jpg\"\n              }\n            ],\n            \"source\": \"3D coordinate system diagram illustrating data rotation and halfspace separation\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center, spanning 80% width\",\n            \"image_style\": \"flat vector graphics\",\n            \"container\": null,\n            \"special_instructions\": \"The image depicts a 3D coordinate system. On the left, a cloud of green dots is shown in a standard orientation. A curved arrow indicates a rotation. On the right, the same data cloud is rotated, and a semi-transparent blue polygon (representing a halfspace) is superimposed, cutting through the data. The axes are drawn in olive-brown.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/d178a4db-6542-48c9-961e-e63d7c4d6299_input_0.jpg"], "target_image": "images/multi_image/d178a4db-6542-48c9-961e-e63d7c4d6299_target.png", "width": 1440, "height": 1080}
+{"index": 138, "id": "959e3975-1c26-4e03-9540-9d31ac2ff802", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#8B0000\",\n      \"secondary\": \"#F5F5DC\",\n      \"text\": \"#000000\",\n      \"accent\": \"#000000\"\n    },\n    \"special_instructions\": \"The slide features a clean, high-contrast academic style. The background is pure white. The primary accent color is a deep, dark red (#8B0000) used for the main title and the header of the right-hand text box. A light beige/cream color (#F5F5DC) is used as a background for the 'Probability functional descent' text box to distinguish it from the white background. The typography is sans-serif, likely Arial or Helvetica, with bold weights used for emphasis on key terms like 'advantage', 'policy', 'influence function', and 'PFD recovers'.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Technical, Educational, Professional\",\n    \"composition_style\": \"Asymmetric, Two-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 12,\n      \"blocks\": [\n        {\n          \"id\": \"title-12\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Reinforcement learning\",\n          \"style\": {\n            \"color\": \"#8B0000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, dark red text positioned at the top left of the slide.\"\n          }\n        },\n        {\n          \"id\": \"equation-12-1\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"J_{RL}(\\\\pi) = \\\\mathbb{E}_{\\\\pi}[\\\\sum_t \\\\gamma^t R_t]\",\n            \"variables\": null\n          },\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"below title, left side\",\n            \"container\": null,\n            \"special_instructions\": \"Standard mathematical notation centered below the title.\"\n          }\n        },\n        {\n          \"id\": \"body-12-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"1. Approximate the advantage Q^\\\\pi(s,a) - V^\\\\pi(s), the influence function for J_RL\\n2. Update the policy \\\\pi\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left side, below equation\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Numbered list. Key terms 'advantage', 'policy' are bolded.\"\n          }\n        },\n        {\n          \"id\": \"body-12-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"PFD recovers:\\n• Policy gradient\\n• Actor-critic\\n• Dual actor critic\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left side, below numbered list\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points using solid black circles. 'PFD recovers:' acts as a sub-header.\"\n          }\n        },\n        {\n          \"id\": \"body-12-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Probability functional descent\\n1. Compute the influence function \\\\Psi of J at \\\\mu\\n2. Choose \\\\mu' such that\\nE_{x \\\\sim \\\\mu'}[\\\\Psi(x)] < E_{x \\\\sim \\\\mu}[\\\\Psi(x)]\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"right side, inside a beige box\",\n            \"container\": \"Beige background box (#F5F5DC)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is contained within a rectangular box with a light beige background. The title 'Probability functional descent' is bold. The mathematical inequality is centered at the bottom of the text block.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_783_424_1424_791.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_783_424_1424_791.jpg\"\n              }\n            ],\n            \"source\": \"Diagram of an Agent interacting with an environment (Atari game), showing State input, DNN processing, Policy output, Reward feedback, and Observe state loop.\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-right\",\n            \"image_style\": \"Line art diagram with a small inset screenshot of a game (likely Breakout) showing the agent's view.\",\n            \"container\": null,\n            \"special_instructions\": \"The diagram illustrates the reinforcement learning loop. It includes a box labeled 'Agent' containing a 'State' input and a 'DNN' (Deep Neural Network) with 'parameter \\\\theta'. Arrows indicate the flow from State to DNN to 'policy \\\\pi_\\\\theta(s, a)', then to 'Take action', which results in 'Reward' and 'Observe state' feedback loops. A small screenshot of a game is embedded in the diagram.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/959e3975-1c26-4e03-9540-9d31ac2ff802_input_0.jpg"], "target_image": "images/multi_image/959e3975-1c26-4e03-9540-9d31ac2ff802_target.png", "width": 1440, "height": 810}
+{"index": 139, "id": "40f72cc9-027a-4054-a787-8360409281df", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#1E3A8A\",\n      \"secondary\": \"#2563EB\",\n      \"text\": \"#1F2937\",\n      \"accent\": \"#FCA5A5\"\n    },\n    \"special_instructions\": \"Clean, high-contrast academic style. The slide relies on a stark white background with dark grey text for readability. Key terms are highlighted in a medium blue. The visual hierarchy is established through font weight and size. The diagram uses pastel blue and pink ovals to distinguish between object nodes and predicate nodes.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Educational, Technical\",\n    \"composition_style\": \"Asymmetric, Top-down\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 3,\n      \"blocks\": [\n        {\n          \"id\": \"title-3\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"WHAT IS SCENE GRAPH GENERATION (SGG) ?\",\n          \"style\": {\n            \"color\": \"#1E3A8A\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, dark blue text aligned to the left margin.\"\n          }\n        },\n        {\n          \"id\": \"body-3-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SGG aims to represent observable knowledges in an image in the form of a graph\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point. The phrase 'observable knowledges in an image' is highlighted in blue (#2563EB).\"\n          }\n        },\n        {\n          \"id\": \"body-3-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"The knowledge includes 1) object information and 2) their relation information, which is mapped to a scene graph\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, below previous body text\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point. The phrase 'object information' and 'their relation information' are highlighted in blue (#2563EB).\"\n          }\n        },\n        {\n          \"id\": \"body-3-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"E.g., Object information: {man, horse, glasses, bucket}\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"minor\",\n            \"position\": \"top-left, indented\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Sub-bullet point. The list of objects is enclosed in curly braces.\"\n          }\n        },\n        {\n          \"id\": \"body-3-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"E.g., Relationship information between objects: {feeding, wearing, ..., holding, eat from}\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"minor\",\n            \"position\": \"top-left, indented\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Sub-bullet point. The list of relationships is enclosed in curly braces.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_585_925_1337_1501.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_585_925_1337_1501.jpg\"\n              }\n            ],\n            \"source\": \"Input image showing a man feeding a horse with bounding boxes\",\n            \"caption\": \"Scene Graph Generation by Iterative Message Passing. Danfei Xu et al. CVPR'2017\"\n          },\n          \"style\": {\n            \"position\": \"bottom-left\",\n            \"image_style\": \"Standard photograph with colored bounding boxes overlaid\",\n            \"container\": null,\n            \"special_instructions\": \"The image has text labels 'Input (image)' above it and 'scene graph generation' vertically to its left. The caption is located below the image.\"\n          }\n        },\n        {\n          \"id\": \"caption-3-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Scene Graph Generation by Iterative Message Passing. Danfei Xu et al. CVPR'2017\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1F2937\",\n            \"position\": \"below figure-1\",\n            \"text_style\": null,\n            \"special_instructions\": null\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1362_924_2230_1567.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1362_924_2230_1567.jpg\"\n              }\n            ],\n            \"source\": \"Scene graph diagram output\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-right\",\n            \"image_style\": \"Conceptual diagram with nodes and arrows\",\n            \"container\": null,\n            \"special_instructions\": \"The diagram shows a flow from 'SGG Model' to a graph structure. It includes a legend at the bottom distinguishing 'object' (blue oval) and 'predicate' (pink oval).\"\n          }\n        },\n        {\n          \"id\": \"illustration-3-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large, light grey arrow pointing from left to right labeled 'SGG Model', indicating the transformation process from input to output.\",\n          \"style\": {\n            \"illustration_style\": \"Flowchart connector\",\n            \"color_scheme\": \"Light grey fill with black outline\",\n            \"position\": \"between the input image and the output graph\",\n            \"special_instructions\": \"The arrow is thick and blocky, serving as a visual bridge between the two main visual elements.\"\n          }\n        },\n        {\n          \"id\": \"illustration-3-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A legend consisting of two colored ovals with text labels: a light blue oval labeled 'object' and a light pink oval labeled 'predicate'.\",\n          \"style\": {\n            \"illustration_style\": \"Legend key\",\n            \"color_scheme\": \"Light blue (#BFDBFE) and Light pink (#FECACA)\",\n            \"position\": \"bottom-right, below the scene graph diagram\",\n            \"special_instructions\": \"The ovals are horizontally aligned with their respective text labels.\"\n          }\n        },\n        {\n          \"id\": \"footer-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"3\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right corner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Slide number, small and unobtrusive.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/40f72cc9-027a-4054-a787-8360409281df_input_0.jpg", "images/multi_image/40f72cc9-027a-4054-a787-8360409281df_input_1.jpg"], "target_image": "images/multi_image/40f72cc9-027a-4054-a787-8360409281df_target.png", "width": 2880, "height": 1620}
+{"index": 140, "id": "a3ffabdd-8ad0-49f9-b553-fc4baed218ed", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#1E73BE\",\n      \"secondary\": \"#A30000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#F5E6A3\"\n    },\n    \"special_instructions\": \"The slide features a clean, white background typical of academic or technical presentations. The design relies heavily on color coding to distinguish concepts: Blue for primary data/model, Red for counterfactuals/issues, and Green for the main contribution. There are distinct background color blocks (yellow and light green) used to separate sections like the 'Question' and 'Main Contribution'. The font is a standard sans-serif (likely Arial), ensuring high readability.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Informative\",\n    \"composition_style\": \"Asymmetric, Top-down flow\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 2,\n      \"blocks\": [\n        {\n          \"id\": \"title-2\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Training a surrogate model using all the queried datapoints (y=0/1) and one-sided counterfactuals (for datapoints with y=0)\",\n          \"style\": {\n            \"color\": \"#1E73BE\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large blue text spanning the top width of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-2-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Counterfactuals as ordinary labelled instances?\\nDecision boundary shift issue\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#A30000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Red text, bold weight, positioned to the left of the central diagram.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_874_171_1676_521.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_874_171_1676_521.jpg\"\n              }\n            ],\n            \"source\": \"Diagram illustrating decision boundaries, data points, and counterfactuals\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-right\",\n            \"image_style\": \"Standard plot with black axes and grid lines\",\n            \"container\": null,\n            \"special_instructions\": \"The image contains a scatter plot with blue and orange dots, a black solid line (Target model), a red dashed line (Surrogate model), and black triangles (Counterfactuals). It includes a legend box.\"\n          }\n        },\n        {\n          \"id\": \"body-2-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Question: Can we improve model reconstruction specifically leveraging the fact that the counterfactuals are quite close to the boundary?\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle, spanning width\",\n            \"container\": \"Background color block\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is placed inside a light yellow rectangular background block that spans the width of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-2-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Main Contribution:\\nNovel Model Reconstruction Strategies Using Counterfactuals With Theoretical Guarantees From Polytope Theory\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"bottom-center\",\n            \"container\": \"Rounded rectangular background box\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a light green rounded rectangle. 'Main Contribution:' is in red, while the rest of the text is black.\"\n          }\n        },\n        {\n          \"id\": \"body-2-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Related Works: [Aivodji et al.'20][Wang et al.'22]\\nOther Privacy + CF: [Pawelczyk et al.'23][Goethals et al.'23] [Yadav et al.'23]\\nModel extraction in other settings: [Gong et al.'20] [Milli et al.'19]\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small text at the very bottom left corner, listing citations.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/a3ffabdd-8ad0-49f9-b553-fc4baed218ed_input_0.jpg"], "target_image": "images/multi_image/a3ffabdd-8ad0-49f9-b553-fc4baed218ed_target.png", "width": 1920, "height": 1080}
+{"index": 141, "id": "0a469adc-d057-4d80-ac97-9220594a6afd", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#4F66A5\",\n      \"secondary\": \"#F28C8C\",\n      \"text\": \"#333333\",\n      \"accent\": \"#4F66A5\"\n    },\n    \"special_instructions\": \"Clean, minimalist academic style typical of LaTeX Beamer or PowerPoint templates. High contrast between text and background. Uses a standard sans-serif font family. The chart utilizes a specific palette of light blue and light red for data visualization.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Data-driven\",\n    \"composition_style\": \"Left-aligned title, centered chart, footer bar\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 37,\n      \"blocks\": [\n        {\n          \"id\": \"title-37\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Improved Adversarial Training\",\n          \"style\": {\n            \"color\": \"#4F66A5\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large font size, dark blue color, aligned to the left margin.\"\n          }\n        },\n        {\n          \"id\": \"body-37-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Stronger attacks improve adversarial training!\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style with a small dark blue diamond/bullet marker.\"\n          }\n        },\n        {\n          \"id\": \"body-37-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"adversarial accuracy of models on CIFAR-10 (adversarial training)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"above the chart, centered horizontally\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Chart title text, smaller font size than main body text.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_47_191_674_426.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_47_191_674_426.jpg\"\n              }\n            ],\n            \"source\": \"Bar chart comparing adversarial accuracy\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"centered below the chart title\",\n            \"image_style\": \"Standard bar chart with grid lines\",\n            \"container\": null,\n            \"special_instructions\": \"The chart displays two sets of bars for each epsilon value. Blue bars represent 'Wong et al. (2019)' and red bars represent 'FW + dual LMO (ours)'. The legend is located inside the chart area at the bottom.\"\n          }\n        },\n        {\n          \"id\": \"footer-37\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"K.Wu, A.Wang and Y.Yu | Wasserstein Adversarial Attacks | July 29, 2020 | 17 / 18\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom of the slide\",\n            \"container\": \"Dark blue horizontal bar spanning full width\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"White text on a dark blue background bar. Contains author names, presentation title, date, and slide number.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/0a469adc-d057-4d80-ac97-9220594a6afd_input_0.jpg"], "target_image": "images/multi_image/0a469adc-d057-4d80-ac97-9220594a6afd_target.png", "width": 726, "height": 545}
+{"index": 142, "id": "8520a93d-7f0f-4c3b-95bd-4fc3f5a68b2d", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Research Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#D80000\",\n      \"secondary\": \"#D80000\",\n      \"text\": \"#8B0000\",\n      \"accent\": \"#D80000\"\n    },\n    \"special_instructions\": \"The slide features a clean, white background typical of academic presentations. The design relies heavily on the color red (#D80000) for branding and emphasis. The charts are enclosed in a thin red border, creating a unified visual container for the data. The font choice is a standard sans-serif (likely Arial), prioritizing readability of the scientific data.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Data-Driven\",\n    \"composition_style\": \"Centered, Grid-based\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 8,\n      \"blocks\": [\n        {\n          \"id\": \"title-8\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Experimental Results\",\n          \"style\": {\n            \"color\": \"#D80000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold red text located in the upper-left quadrant of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-8-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"LUI is trustworthy in predicting its own performance\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#8B0000\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Dark red text, slightly smaller than the title, serving as the main thesis statement.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_94_482_996_1164.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_94_482_996_1164.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing 'Betting Game Favourable' results with multiple colored trajectories and shaded confidence intervals.\",\n            \"caption\": \"Betting Game Favourable\"\n          },\n          \"style\": {\n            \"position\": \"center-left\",\n            \"image_style\": \"Thin red border around the plot area\",\n            \"container\": \"Red border box\",\n            \"special_instructions\": \"The chart displays 'Estimation Error' on the Y-axis and 'Trajectory' (log scale) on the X-axis. It features blue, red, and orange lines with corresponding shaded areas.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1009_487_1886_1161.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1009_487_1886_1161.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing 'Betting Game Unfavourable' results.\",\n            \"caption\": \"Betting Game Unfavourable\"\n          },\n          \"style\": {\n            \"position\": \"center\",\n            \"image_style\": \"Thin red border around the plot area\",\n            \"container\": \"Red border box\",\n            \"special_instructions\": \"The chart displays 'Estimation Error' on the Y-axis and 'Trajectory' (log scale) on the X-axis. It features an orange line with a shaded confidence interval and a flat red line at y=0.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1899_484_2789_1164.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1899_484_2789_1164.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing 'Bandit' results.\",\n            \"caption\": \"Bandit\"\n          },\n          \"style\": {\n            \"position\": \"center-right\",\n            \"image_style\": \"Thin red border around the plot area\",\n            \"container\": \"Red border box\",\n            \"special_instructions\": \"The chart displays 'Estimation Error' on the Y-axis and 'Trajectory' (log scale) on the X-axis. It features blue, orange, red, and green lines with shaded confidence intervals, along with horizontal dotted and solid reference lines.\"\n          }\n        },\n        {\n          \"id\": \"footer-8\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Radboud University\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#8B0000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small red text located at the bottom center of the slide, accompanied by a small university logo icon.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/8520a93d-7f0f-4c3b-95bd-4fc3f5a68b2d_input_0.jpg", "images/multi_image/8520a93d-7f0f-4c3b-95bd-4fc3f5a68b2d_input_1.jpg", "images/multi_image/8520a93d-7f0f-4c3b-95bd-4fc3f5a68b2d_input_2.jpg"], "target_image": "images/multi_image/8520a93d-7f0f-4c3b-95bd-4fc3f5a68b2d_target.png", "width": 2880, "height": 1620}
+{"index": 143, "id": "17e37912-5bc0-4c14-a412-ecb52a33bd2e", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#800000\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#800000\"\n    },\n    \"special_instructions\": \"The slide features a clean, high-contrast academic layout. The top header is a solid deep red banner containing a white University of Utah logo on the left and the title in white text. The main content area is white with black text. Mathematical equations and specific terms like 'Specturm bias' are highlighted in red. The footer is a black bar with white text.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Academic\",\n    \"composition_style\": \"Asymmetric, Header-Banner, Two-Column Content\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 3,\n      \"blocks\": [\n        {\n          \"id\": \"title-3\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Hard cases: high-freq. + multi-scale PDEs\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center, inside red banner\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Located inside a solid deep red (#800000) header bar. The text is white and bold.\"\n          }\n        },\n        {\n          \"id\": \"body-3-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Examples:\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, below header\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text. Standard black font.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_265_161_600_400.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_265_161_600_400.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing a high-frequency oscillating function\",\n            \"caption\": \"sin(500x) - 2(x - 0.5)^2\"\n          },\n          \"style\": {\n            \"position\": \"top-left, next to the 'Examples:' text\",\n            \"image_style\": \"Standard plot with axes and grid lines\",\n            \"container\": null,\n            \"special_instructions\": \"A line chart with a red oscillating line. The caption is located directly below the chart.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_846_140_1103_341.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_846_140_1103_341.jpg\"\n              }\n            ],\n            \"source\": \"Heatmap image showing a multi-scale pattern\",\n            \"caption\": \"(sin(x) + 0.1 sin(20x) + cos(100x)) (sin(y) + 0.1 sin(20y) + cos(100y))\"\n          },\n          \"style\": {\n            \"position\": \"top-right, aligned with the first figure\",\n            \"image_style\": \"Heatmap with a colorbar at the top\",\n            \"container\": null,\n            \"special_instructions\": \"A square heatmap image showing a grid-like pattern of red and yellow dots. A colorbar is visible above the image. The caption is located directly below the image.\"\n          }\n        },\n        {\n          \"id\": \"body-3-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Current NN-based ML solvers hard to handle such cases, because:\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left, below the figures\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text. Standard black font.\"\n          }\n        },\n        {\n          \"id\": \"body-3-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"- “Specturm bias” [1] in NN training\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left, indented under previous point\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text. The term 'Specturm bias' is highlighted in red.\"\n          }\n        },\n        {\n          \"id\": \"body-3-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"- Easy to capture low-freq. info, hard to capture high-freq.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left, indented under previous point\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text. Standard black font.\"\n          }\n        },\n        {\n          \"id\": \"footer-3\",\n          \"type\": \"text\",\n          \"role\": \"footer\",\n          \"content\": \"3 | [1] Rahman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F., Bengio, Y., and Courville, A.(2019). On the spectral bias of neural networks. ICML\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom, spanning full width\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Located inside a black footer bar. Slide number '3' is on the left, citation text is on the right.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/17e37912-5bc0-4c14-a412-ecb52a33bd2e_input_0.jpg", "images/multi_image/17e37912-5bc0-4c14-a412-ecb52a33bd2e_input_1.jpg"], "target_image": "images/multi_image/17e37912-5bc0-4c14-a412-ecb52a33bd2e_target.png", "width": 1440, "height": 810}
+{"index": 144, "id": "e2f5826a-a60f-4d7b-b1b2-e32c7bf36ff1", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#333333\",\n      \"secondary\": \"#666666\",\n      \"text\": \"#000000\",\n      \"accent\": \"#E67E22\"\n    },\n    \"special_instructions\": \"Clean, minimalist academic style. High contrast black text on white background. Uses standard sans-serif fonts. The diagram utilizes specific colors (orange, green, blue) to denote different sets of views.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Two-column layout with title centered at top\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 14,\n      \"blocks\": [\n        {\n          \"id\": \"title-14\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Self-Supervised Training\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, centered at the very top of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-14-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Learning Geometry in Occluded Regions\\n\\nTraditional reprojection loss formulations do not give training signals for areas occluded in the input image.\\n\\n→ Our density field allows reconstructing any frame from any other frame\\n\\n→ We can reconstruct P in view I2 by sampling colors from I3\\n\\n→ To minimize the loss, our network has to predict correct geometry for P, even though P is occluded in I1\\n\\n→ This requires at least two extra views other than the input view.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"right half, vertically centered\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points represented by right-pointing arrows (→). Text is left-aligned within the right column.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_90_350_798_722.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_90_350_798_722.jpg\"\n              }\n            ],\n            \"source\": \"Diagram illustrating self-supervised training geometry\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left half, vertically centered\",\n            \"image_style\": \"Schematic diagram with geometric shapes (triangles and squares) and connecting lines. Includes a legend box at the bottom.\",\n            \"container\": null,\n            \"special_instructions\": \"The diagram features a red oval circling views I2 and I3, a red circle around patch P, and red crosses indicating occlusion. Green checkmarks indicate successful reconstruction.\"\n          }\n        },\n        {\n          \"id\": \"equation-14-1\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"$$ \\\\mathbf{I}\\\\in N_{\\\\mathrm{render}} $$\",\n            \"variables\": \"I: Image, N_render: Set of rendering views\"\n          },\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"left side, below the diagram\",\n            \"container\": null,\n            \"special_instructions\": \"Part of the legend box in the diagram.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/e2f5826a-a60f-4d7b-b1b2-e32c7bf36ff1_input_0.jpg"], "target_image": "images/multi_image/e2f5826a-a60f-4d7b-b1b2-e32c7bf36ff1_target.png", "width": 1920, "height": 1080}
+{"index": 145, "id": "235841bd-9bda-4af3-8caf-bf09939db2df", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Helvetica Neue\",\n      \"secondary\": \"Helvetica Neue\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#A83F39\",\n      \"secondary\": \"#D45D79\",\n      \"text\": \"#1A1A1A\",\n      \"accent\": \"#E6B800\"\n    },\n    \"special_instructions\": \"Clean, minimalist academic style typical of university lectures. High contrast black text on white background. Uses a sans-serif font family (likely Helvetica or Arial). The layout is grid-based with distinct sections for title, visual evidence, data plots, and a legend. Colors are used functionally to distinguish data series in the charts and legend items.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Informative\",\n    \"composition_style\": \"Asymmetric Grid\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 76,\n      \"blocks\": [\n        {\n          \"id\": \"title-76\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"IDRL can learn from comparisons of trajectories\",\n          \"style\": {\n            \"color\": \"#1A1A1A\",\n            \"font_family\": \"Helvetica Neue\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold sans-serif font. Positioned at the top left, spanning roughly 40% of the slide width.\"\n          }\n        },\n        {\n          \"id\": \"body-76-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Environment by Sadigh et al. 2017\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left, below the environment image\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small, regular weight text serving as a caption for the image above it.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_72_268_634_1078.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_72_268_634_1078.jpg\"\n              }\n            ],\n            \"source\": \"Simulation environment showing a car trajectory on a road\",\n            \"caption\": \"Environment by Sadigh et al. 2017\"\n          },\n          \"style\": {\n            \"position\": \"left column, vertically centered\",\n            \"image_style\": \"No border or shadow\",\n            \"container\": null,\n            \"special_instructions\": \"Shows a top-down view of a road with a car and trajectory lines.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_669_274_1674_1140.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_669_274_1674_1140.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing 'estimated regret' vs 'number of queries'\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center column\",\n            \"image_style\": \"Standard matplotlib plot style with black axes and grid lines\",\n            \"container\": null,\n            \"special_instructions\": \"X-axis: number of queries (0-500). Y-axis: estimated regret (0-2). Shows multiple colored lines starting high and dropping near 0.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1678_265_2733_1162.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1678_265_2733_1162.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing 'cos-similarity' vs 'number of queries'\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column\",\n            \"image_style\": \"Standard matplotlib plot style with black axes\",\n            \"container\": null,\n            \"special_instructions\": \"X-axis: number of queries (0-500). Y-axis: cos-similarity (0.0-0.4). Shows lines starting near 0 and rising to ~0.4.\"\n          }\n        },\n        {\n          \"id\": \"body-76-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Information Directed Reward Learning (ours)\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left, part of the legend\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Part of the legend group. Text is black.\"\n          }\n        },\n        {\n          \"id\": \"body-76-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Maximum Regret (Wilde et al. 2020)\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left, below the first legend item\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Part of the legend group. Text is black.\"\n          }\n        },\n        {\n          \"id\": \"body-76-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Expected Policy Divergence (EPD, adapted from Daniel et al. 2015)\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-center, part of the legend\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Part of the legend group. Text is black.\"\n          }\n        },\n        {\n          \"id\": \"body-76-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Information Gain on the Reward\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right, part of the legend\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Part of the legend group. Text is black.\"\n          }\n        },\n        {\n          \"id\": \"illustration-76-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A small rectangular box with a dark red background and white text located in the top right corner.\",\n          \"style\": {\n            \"illustration_style\": \"Label box\",\n            \"color_scheme\": \"Dark red background (#A83F39), white text\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"Contains the text 'GP reward model'.\"\n          }\n        },\n        {\n          \"id\": \"caption-76-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"12\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#1A1A1A\",\n            \"position\": \"bottom-right corner\",\n            \"text_style\": null,\n            \"special_instructions\": \"Slide number.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/235841bd-9bda-4af3-8caf-bf09939db2df_input_0.jpg", "images/multi_image/235841bd-9bda-4af3-8caf-bf09939db2df_input_1.jpg", "images/multi_image/235841bd-9bda-4af3-8caf-bf09939db2df_input_2.jpg"], "target_image": "images/multi_image/235841bd-9bda-4af3-8caf-bf09939db2df_target.png", "width": 2880, "height": 1620}
+{"index": 146, "id": "e6fd43ff-4fe5-47b8-9480-ff45a29d4edd", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#D62828\",\n      \"secondary\": \"#555555\",\n      \"text\": \"#555555\",\n      \"accent\": \"#D62828\"\n    },\n    \"special_instructions\": \"Clean, white background with a distinct red footer bar. The design is strictly academic and technical, utilizing standard sans-serif fonts. Visual hierarchy is established through font weight and color (red for titles and key mathematical terms).\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Technical, Professional, Academic\",\n    \"composition_style\": \"Two-column Asymmetric\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 4,\n      \"blocks\": [\n        {\n          \"id\": \"title-4\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"HETEROGENEOUS MOTIF GRAPH\",\n          \"style\": {\n            \"color\": \"#D62828\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, red text aligned to the left margin.\"\n          }\n        },\n        {\n          \"id\": \"body-4-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Based on the motif vocabulary, we build a heterogeneous graph that contains motif nodes and molecular nodes.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#555555\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard body text, grey color, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_103_258_788_725.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_103_258_788_725.jpg\"\n              }\n            ],\n            \"source\": \"Heterogeneous graph diagram showing molecular nodes connected to motif nodes\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left-center\",\n            \"image_style\": \"None\",\n            \"container\": null,\n            \"special_instructions\": \"Complex diagram containing chemical structures, nodes, and edges. Includes a legend for Atom Type, Bond Type, and Molecular Type.\"\n          }\n        },\n        {\n          \"id\": \"equation-4-1\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"A_{ij} = \\\\begin{cases} PMI_{ij}, & \\\\text{if } i, j \\\\text{ are motifs} \\\\\\\\ TF-IDF_{ij}, & \\\\text{if } i \\\\text{ or } j \\\\text{ is a motif} \\\\\\\\ 0, & \\\\text{Otherwise} \\\\end{cases}\",\n            \"variables\": \"A_ij (Edge weight), PMI (Pointwise Mutual Information), TF-IDF (Term Frequency-Inverse Document Frequency)\"\n          },\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#D62828\",\n            \"position\": \"right-center\",\n            \"container\": null,\n            \"special_instructions\": \"Mathematical definition of edge weight. The term 'Edge weight:' is in red, followed by the piecewise function in black/red math font.\"\n          }\n        },\n        {\n          \"id\": \"equation-4-2\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"TF-IDF_{ij} = C(i)_j \\\\left( \\\\log \\\\frac{1+M}{1+N(i)} + 1 \\\\right)\",\n            \"variables\": \"TF-IDF_ij, C(i)_j, M, N(i)\"\n          },\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#D62828\",\n            \"position\": \"right-center\",\n            \"container\": null,\n            \"special_instructions\": \"Formula for calculating TF-IDF. Text color is red.\"\n          }\n        },\n        {\n          \"id\": \"equation-4-3\",\n          \"type\": \"text\",\n          \"role\": \"equation\",\n          \"content\": {\n            \"latex\": \"PMI_{ij} = \\\\log \\\\frac{p(i,j)}{p(i)p(j)}\",\n            \"variables\": \"PMI_ij, p(i,j), p(i), p(j)\"\n          },\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#D62828\",\n            \"position\": \"right-center\",\n            \"container\": null,\n            \"special_instructions\": \"Formula for calculating PMI. Text color is red.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/e6fd43ff-4fe5-47b8-9480-ff45a29d4edd_input_0.jpg"], "target_image": "images/multi_image/e6fd43ff-4fe5-47b8-9480-ff45a29d4edd_target.png", "width": 1440, "height": 1080}
+{"index": 147, "id": "9ae4435b-5090-43f1-88c2-c0faafdb2f0e", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Report\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#005691\",\n      \"secondary\": \"#E0E0E0\",\n      \"text\": \"#000000\",\n      \"accent\": \"#D93025\"\n    },\n    \"special_instructions\": \"The slide features a clean, professional academic layout. The top section has a solid dark blue header bar containing the title and university logo. The main content area is white with a large, complex data table on the left and a vertical list of model descriptions on the right. The table uses a standard grid layout with specific color coding for statistical significance (Red for best, Blue for second-best). The overall feel is strictly functional and data-centric.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Data-Heavy\",\n    \"composition_style\": \"Two-column layout with a header bar\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 11,\n      \"blocks\": [\n        {\n          \"id\": \"title-11\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Quantitative evaluation\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Located inside a solid dark blue (#005691) header bar that spans the full width of the slide.\"\n          }\n        },\n        {\n          \"id\": \"logo-11\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [],\n            \"source\": \"Tongji University Logo\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-right\",\n            \"image_style\": \"Vector logo with text 'TONGJI UNIVERSITY' below the emblem\",\n            \"container\": null,\n            \"special_instructions\": \"White logo placed inside the dark blue header bar on the right side.\"\n          }\n        },\n        {\n          \"id\": \"caption-11-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Table 2: Quantitative results. The best and second-best results are highlighted by red and blue. ★ refers to the inference results of this method are not affected by seed. † means the inference results of this method are affected by the seed. We tested the results five times by varying the seed, taking the average and standard deviation.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"top-left, below header\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"special_instructions\": \"Small text block describing the table metrics and symbols.\"\n          }\n        },\n        {\n          \"id\": \"table-11\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_table_63_270_1139_993.jpg\",\n                \"relative_path\": \"imgs/img_in_table_63_270_1139_993.jpg\"\n              }\n            ],\n            \"source\": \"Quantitative Results Table\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left-column, vertically centered\",\n            \"image_style\": \"Standard academic table with grid lines. Contains numerical data with standard deviation rows. Specific cells are colored red (#D93025) and blue (#005691) to highlight top results.\",\n            \"container\": null,\n            \"special_instructions\": \"The table is split into two main sections: 'UDIS - D_test' and 'UDIS - D_train'. The bottom section is labeled 'SRStitcher Variants' with a light grey background header.\"\n          }\n        },\n        {\n          \"id\": \"body-11-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SRStitcher-S\\n\\nbased on\\n\\nStable-Diffusion-2 model\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"right-column, top\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Vertical text block describing the first model variant. Bold text for the model name, normal text for the description.\"\n          }\n        },\n        {\n          \"id\": \"body-11-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SRStitcher-U\\n\\nbased on\\n\\nStable-Diffusion-2-1-Unclip model\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"right-column, middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Vertical text block describing the second model variant.\"\n          }\n        },\n        {\n          \"id\": \"body-11-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SRStitcher-C\\n\\nbased on\\n\\nStable-Diffusion-control-v11p-sd15-inpaint model\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"right-column, bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Vertical text block describing the third model variant.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/9ae4435b-5090-43f1-88c2-c0faafdb2f0e_input_0.jpg"], "target_image": "images/multi_image/9ae4435b-5090-43f1-88c2-c0faafdb2f0e_target.png", "width": 1920, "height": 1080}
+{"index": 148, "id": "c463644b-f771-48b6-9913-bde1f81df318", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Minimalist\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#000000\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#D63031\"\n    },\n    \"special_instructions\": \"The slide features a stark, high-contrast academic design. The background is pure white (#FFFFFF), providing maximum clarity for the black text and diagrams. The visual hierarchy relies entirely on font weight and size rather than color. There are no decorative backgrounds, gradients, or textures. The layout is clean and functional, typical of academic conference slides.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Asymmetric, Top-Heavy\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 2,\n      \"blocks\": [\n        {\n          \"id\": \"title-2\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Model-based RL\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold sans-serif font. Positioned in the top-left corner with significant whitespace to the right.\"\n          }\n        },\n        {\n          \"id\": \"body-2-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"In model-based RL, the agent actively collects data from the environment and learns an inevitably inaccurate model of it. To maximize the utility, the agent faces the dilemma of exploration and exploitation and trades off between\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard paragraph text. The phrase 'dilemma of exploration and exploitation' is bolded for emphasis.\"\n          }\n        },\n        {\n          \"id\": \"body-2-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• exploring the uncertain part of the environment to get information and\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center, below paragraph\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point. The phrase 'exploring the uncertain part of the environment' is bolded.\"\n          }\n        },\n        {\n          \"id\": \"body-2-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• exploiting the existing observations to make decisions greedily.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center, below previous bullet\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point. The phrase 'exploiting the existing observations' is bolded.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_314_717_977_1195.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_314_717_977_1195.jpg\"\n              }\n            ],\n            \"source\": \"Conceptual cycle diagram showing Policy -> Data -> Model loop\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-left\",\n            \"image_style\": \"Line art diagram with black arrows and text labels on white background\",\n            \"container\": null,\n            \"special_instructions\": \"A circular flow diagram illustrating the Model-based RL cycle. It contains four nodes: 'policy', 'data', 'model', and labels for 'acting', 'fitting', and 'updating'. The text is bold and sans-serif.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1258_732_1871_1200.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1258_732_1871_1200.jpg\"\n              }\n            ],\n            \"source\": \"Chart showing Prediction vs Ground Truth with confidence intervals\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-right\",\n            \"image_style\": \"2D line chart with axes, legend, and shaded confidence interval\",\n            \"container\": null,\n            \"special_instructions\": \"A chart plotting 'Reward' vs 'Action'. It features a red dashed line (Ground truth), a black solid line (Prediction), red dots (Observations), and a grey shaded area representing the 95% confidence interval. A legend is present in the top-left corner of the chart area.\"\n          }\n        },\n        {\n          \"id\": \"caption-2-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"36th Conference on Neural Information Processing Systems (NeurIPS 2022)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"bottom-center\",\n            \"text_style\": null,\n            \"special_instructions\": \"Small footer text indicating the conference source.\"\n          }\n        },\n        {\n          \"id\": \"illustration-2-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Two logos in the top-right corner: an orange and blue 'LAMDA' logo and a purple shield-shaped logo for 'Nanjing University'.\",\n          \"style\": {\n            \"illustration_style\": \"Brand logos\",\n            \"color_scheme\": \"Orange, Blue, Purple, White\",\n            \"position\": \"top-right\",\n            \"special_instructions\": \"The LAMDA logo features orange and blue text with a URL below it. The Nanjing University logo is a purple shield with white text and symbols.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/c463644b-f771-48b6-9913-bde1f81df318_input_0.jpg", "images/multi_image/c463644b-f771-48b6-9913-bde1f81df318_input_1.jpg"], "target_image": "images/multi_image/c463644b-f771-48b6-9913-bde1f81df318_target.png", "width": 2304, "height": 1296}
+{"index": 149, "id": "adfe0cd9-63d2-4792-a317-f07c04a7baf3", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#000000\",\n      \"secondary\": \"#0000FF\",\n      \"text\": \"#000000\",\n      \"accent\": \"#0000FF\"\n    },\n    \"special_instructions\": \"Clean, high-contrast academic style. White background with black text. Blue is used for emphasis (headers, bullet points, arrows). The layout is structured with a header, a text block, and a wide diagram spanning the bottom.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Header-Body-Image\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 6,\n      \"blocks\": [\n        {\n          \"id\": \"title-6\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Part 3: Model\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold text. Positioned at the top left. There is a small logo to the left of the text.\"\n          }\n        },\n        {\n          \"id\": \"logo-6\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A circular logo containing a stylized lamp or torch icon and text, located to the left of the main title.\",\n          \"style\": {\n            \"illustration_style\": \"logo\",\n            \"color_scheme\": \"blue and white\",\n            \"position\": \"top-left corner, adjacent to title\",\n            \"special_instructions\": \"Circular border, contains text 'NEURAL INFORMATION PROCESSING SYSTEMS'.\"\n          }\n        },\n        {\n          \"id\": \"logo-6-right\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A circular logo containing a stylized lamp or torch icon and text, located to the left of the main title.\",\n          \"style\": {\n            \"illustration_style\": \"logo\",\n            \"color_scheme\": \"blue and white\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"Circular border, contains text 'NEURAL INFORMATION PROCESSING SYSTEMS'.\"\n          }\n        },\n        {\n          \"id\": \"body-6-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Input module:\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#0000FF\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Blue text. Preceded by a blue right-pointing triangle bullet point.\"\n          }\n        },\n        {\n          \"id\": \"body-6-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"In this module, we first rearrange the sequence from 1D to 2D based on its natural period. Meanwhile, we employ an adaptive learning approach to determine whether to perform simple normalization.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"below 'Input module:', left-aligned\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Black text. Preceded by a blue right-pointing triangle bullet point.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_36_764_2802_1389.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_36_764_2802_1389.jpg\"\n              }\n            ],\n            \"source\": \"Diagram showing information transmission rearrangement and water-wave process\",\n            \"caption\": \"(a) Information transmission before and after the input sequence rearrangement\\n(b) The water-wave information transmission process\"\n          },\n          \"style\": {\n            \"position\": \"bottom, spanning full width\",\n            \"image_style\": \"diagram with multiple sub-panels\",\n            \"container\": null,\n            \"special_instructions\": \"The image contains three distinct panels labeled (a) and (b). It uses various arrow styles (dashed red, solid red, solid blue, dashed purple) to denote different types of information transmission.\"\n          }\n        },\n        {\n          \"id\": \"caption-6-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"(a) Information transmission before and after the input sequence rearrangement\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"below the left and middle panels of the diagram\",\n            \"text_style\": null,\n            \"special_instructions\": null\n          }\n        },\n        {\n          \"id\": \"caption-6-2\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"(b) The water-wave information transmission process\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"below the right panel of the diagram\",\n            \"text_style\": null,\n            \"special_instructions\": null\n          }\n        },\n        {\n          \"id\": \"footer-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"6\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Page number, small font size, centered at the very bottom.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/adfe0cd9-63d2-4792-a317-f07c04a7baf3_input_0.jpg"], "target_image": "images/multi_image/adfe0cd9-63d2-4792-a317-f07c04a7baf3_target.png", "width": 2880, "height": 1620}
+{"index": 150, "id": "9d98b612-3a26-4cf3-b531-19bdc6e53809", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic/Corporate Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#C0C0C0\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#FFFF00\"\n    },\n    \"special_instructions\": \"The slide features a clean, high-contrast white background typical of academic or corporate presentations. The design relies heavily on bold typography and color coding (yellow highlight, purple numbering) to organize information. A distinct visual element is the 'Cont...' indicator in a light grey box at the top, suggesting this is a continuation of a previous slide. The layout is text-heavy with a decorative graphic in the top right corner.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Educational, Informative\",\n    \"composition_style\": \"Left-aligned text with top-right decorative graphic\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 12,\n      \"blocks\": [\n        {\n          \"id\": \"title-12\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Cont...\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"panel\",\n            \"special_instructions\": \"Text is centered inside a light grey rectangular background box (approx #D3D3D3) with rounded corners.\"\n          }\n        },\n        {\n          \"id\": \"illustration-12-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A stylized graphic resembling a hand-drawn sketch on a whiteboard. It features a black marker pen in the top left corner. The text 'the power of communication' is written in a bold, brush-stroke font, underlined by two parallel lines. Below this, five colored arrows (blue, red, blue, green, red) point downwards to the words 'relationship', 'value', 'customer', 'success', and 'loyalty' respectively.\",\n          \"style\": {\n            \"illustration_style\": \"hand-drawn sketch\",\n            \"color_scheme\": \"black text with multi-colored arrows (blue, red, green) and text\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"The graphic serves as a decorative header for the topic of communication.\"\n          }\n        },\n        {\n          \"id\": \"body-12-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"7. COMMUNICATION\\n\\nIt binds everything together, starting from foundation to roof of TQM.\\n\\nIt acts as a vital link between all elements of TQM.\\n\\nCommunication means a common understanding of ideas between the sender and receiver.\\n\\nThe success of TQM demands communication with and among all the organizational members, suppliers and customers.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The heading '7. COMMUNICATION' is highlighted with a bright yellow background (#FFFF00). The body text is a bulleted list with square bullets.\"\n          }\n        },\n        {\n          \"id\": \"body-12-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"There are different ways of communication\\n\\n1. Downward communication: This is the dominant form of communication in an organization, in this flow of information takes place from the management to the employees.\\n\\n2. Upward communication: By this lower level of employees are able to provide suggestions to upper management, in this flow of information takes place from the employees to the top level management.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The section header 'There are different ways of communication' is in bold. The numbered list items '1. Downward communication' and '2. Upward communication' are highlighted with a light purple background (#E6E6FA).\"\n          }\n        },\n        {\n          \"id\": \"caption-12-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"24/08/2025\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#666666\",\n            \"position\": \"bottom-left\",\n            \"text_style\": null,\n            \"special_instructions\": \"Date displayed in small, grey text.\"\n          }\n        },\n        {\n          \"id\": \"caption-12-2\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"12\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#666666\",\n            \"position\": \"bottom-right\",\n            \"text_style\": null,\n            \"special_instructions\": \"Slide number displayed in small, grey text.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/9d98b612-3a26-4cf3-b531-19bdc6e53809_input_0.jpg"], "target_image": "images/multi_image/9d98b612-3a26-4cf3-b531-19bdc6e53809_target.jpg", "width": 2048, "height": 1152}
+{"index": 151, "id": "2e77f3cb-2241-408c-9aba-937a1352dbf8", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#4285F4\",\n      \"secondary\": \"#D3D3D3\",\n      \"text\": \"#000000\",\n      \"accent\": \"#4285F4\"\n    },\n    \"special_instructions\": \"Clean, high-contrast technical slide. Uses a distinct blue (#4285F4) for headers and footer separators. The diagram uses a monochrome palette with light grey backgrounds for nodes. Typography is sans-serif and highly legible.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Asymmetric Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 532,\n      \"blocks\": [\n        {\n          \"id\": \"title-532\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Inheritance\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold text at the very top left corner.\"\n          }\n        },\n        {\n          \"id\": \"body-532-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Class hierarchy\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Sub-header text, normal weight.\"\n          }\n        },\n        {\n          \"id\": \"body-532-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Classes:\\n• Person\\n• Employee is derived from Person\\n• Student is derived from Person\\n• Faculty is derived from Employee\\n• Staff is derived from Employee\\n• Administrator is derived from Employee\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below sub-header\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point list defining the class relationships.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_923_176_1967_923.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_923_176_1967_923.jpg\"\n              }\n            ],\n            \"source\": \"Class hierarchy diagram\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, vertically centered\",\n            \"image_style\": \"UML-style inheritance diagram with grey boxes and black lines\",\n            \"container\": null,\n            \"special_instructions\": \"Diagram shows 'Person' at the top, branching to 'Employee' and 'Student'. 'Employee' branches to 'Faculty', 'Staff', and 'Administrator'. Triangles indicate inheritance.\"\n          }\n        },\n        {\n          \"id\": \"footer-532-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"BEGINNING C++ PROGRAMMING\\nTerminology and Notation\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-left corner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small footer text indicating the course title and topic.\"\n          }\n        },\n        {\n          \"id\": \"footer-532-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"{LP}LearnProgramming\\nacademy\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#4285F4\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Logo text, blue color matching the theme accent.\"\n          }\n        },\n        {\n          \"id\": \"footer-532-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"INHERITANCE\\nS15-CPP-0060-5\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right corner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Topic and slide ID code.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/2e77f3cb-2241-408c-9aba-937a1352dbf8_input_0.jpg"], "target_image": "images/multi_image/2e77f3cb-2241-408c-9aba-937a1352dbf8_target.jpg", "width": 2048, "height": 1178}
+{"index": 152, "id": "577b29a2-f232-46c3-ac0d-2f9f58b86141", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Financial Analysis\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#003366\",\n      \"secondary\": \"#4A7BA7\",\n      \"text\": \"#555555\",\n      \"accent\": \"#D6E4F5\"\n    },\n    \"special_instructions\": \"The slide utilizes a clean, professional aesthetic typical of financial or strategic consulting presentations. The design relies on a strict grid layout with high contrast between the deep blue header and the white content area. Visual emphasis is achieved through the use of specific shades of blue for data visualization (lines and markers) and light blue for highlighting key growth metrics within the charts.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Analytical, Data-Driven\",\n    \"composition_style\": \"Grid-based, Two-column layout for charts\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 193,\n      \"blocks\": [\n        {\n          \"id\": \"title-193\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"AI Monetization – Foundation Models = OpenAI Revenue +1,050% Annually to $3.7B\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Located within a solid dark blue rectangular banner spanning the top of the slide. The text is white and left-aligned within the banner.\"\n          }\n        },\n        {\n          \"id\": \"body-193-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"ChatGPT Paid Subscribers (MM) & Revenue ($B) – 10/22-4/25, per OpenAI & The Information\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#555555\",\n            \"importance\": \"supporting\",\n            \"position\": \"centered horizontally, below the title banner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Centered text block acting as a subtitle or chart description. Font size is smaller than the title but larger than the axis labels.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_59_531_981_1374.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_59_531_981_1374.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing ChatGPT Paid Subscribers growth\",\n            \"caption\": \"Paid Subscribers\"\n          },\n          \"style\": {\n            \"position\": \"left column, vertically centered\",\n            \"image_style\": \"Standard chart styling with dashed grid lines\",\n            \"container\": null,\n            \"special_instructions\": \"The chart features a solid blue line representing actual data and a dashed blue line representing a trend or projection. A light blue oval annotation highlights a +153% / Year growth rate.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1031_529_2002_1373.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1031_529_2002_1373.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing OpenAI Revenue growth\",\n            \"caption\": \"Revenue\"\n          },\n          \"style\": {\n            \"position\": \"right column, vertically centered\",\n            \"image_style\": \"Standard chart styling with dashed grid lines\",\n            \"container\": null,\n            \"special_instructions\": \"The chart features a solid blue line representing actual data and a dashed blue line representing a trend or projection. A light blue oval annotation highlights a 1,050% / Year growth rate.\"\n          }\n        },\n        {\n          \"id\": \"caption-193-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Source: OpenAI disclosures (as of 4/25), The Information (4/25) (link, link, link & link)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#999999\",\n            \"position\": \"bottom-left, below the charts\",\n            \"text_style\": null,\n            \"special_instructions\": \"Small, grey text providing source attribution for the data presented in the charts.\"\n          }\n        },\n        {\n          \"id\": \"footer-193-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"BOND\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-left corner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Logo or brand name in the footer.\"\n          }\n        },\n        {\n          \"id\": \"footer-193-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"AI – New Entrants = Rapid Revenue Growth\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Slide theme or section header in the footer.\"\n          }\n        },\n        {\n          \"id\": \"footer-193-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"192\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right corner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Slide number in the footer.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/577b29a2-f232-46c3-ac0d-2f9f58b86141_input_0.jpg", "images/multi_image/577b29a2-f232-46c3-ac0d-2f9f58b86141_input_1.jpg"], "target_image": "images/multi_image/577b29a2-f232-46c3-ac0d-2f9f58b86141_target.jpg", "width": 2048, "height": 1536}
+{"index": 153, "id": "a148d73d-3373-4db1-a250-c4ea12e4bba1", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Nature & Sustainability\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#1E5A9E\",\n      \"secondary\": \"#E0F2F7\",\n      \"text\": \"#555555\",\n      \"accent\": \"#1E5A9E\"\n    },\n    \"special_instructions\": \"The slide features a clean, corporate aesthetic with a strong emphasis on nature. The design utilizes a specific shade of deep blue (#1E5A9E) as the primary brand color, used for titles, bullet points, and decorative elements. A subtle, wavy blue pattern runs along the very bottom edge of the slide, mimicking water or waves, reinforcing the 'nature' theme. The layout is a standard two-column grid with a header navigation bar at the top.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Corporate, Eco-friendly\",\n    \"composition_style\": \"Two-column grid with header navigation\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 17,\n      \"blocks\": [\n        {\n          \"id\": \"nav-17\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"01. Desing your work, design your life | 02. Enjoy the Planet | 03. Building your career | 04. Boarding information\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#999999\",\n            \"importance\": \"minor\",\n            \"position\": \"Top center, spanning full width\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Navigation bar text. '02. Enjoy the Planet' is underlined in blue to indicate the current section.\"\n          }\n        },\n        {\n          \"id\": \"title-17\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Support | Living the Values\",\n          \"style\": {\n            \"color\": \"#1E5A9E\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"Top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold blue text. A solid blue square bullet point precedes the word 'Support'.\"\n          }\n        },\n        {\n          \"id\": \"body-17-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Shizen Energy Holiday\\n\\nA Shizen Energy Holiday is a unique paid leave system that can be taken separately from annual paid leave. Eligible crew members receive seven additional days per year to step away from daily work and invest in themselves, with themes of \\\"Mind,\\\" \\\"Skill,\\\" \\\"Body,\\\" or \\\"Enjoying the Planet.\\\"\\n\\nEveryone enjoys their days off in their own way, from outdoor sports, family camping, and volunteer activities, to studying for new certifications or traveling to explore different parts of the Earth.\\n\\nAfter fully enjoying their time off, our crew share their experiences and what they learned with colleagues across the company in a short report.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#555555\",\n            \"importance\": \"dominant\",\n            \"position\": \"Left column, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard body text. The first line 'Shizen Energy Holiday' is bolded and colored blue (#1E5A9E) to act as a sub-heading.\"\n          }\n        },\n        {\n          \"id\": \"body-17-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SANU 2nd Home, shared vacation house service\\n\\nAs a part of our benefits scheme, we are members of SANU 2nd Home, a shared vacation home concept based on the idea of \\\"another home in nature.\\\"\\n\\nThe facilities feature environmentally-friendly architecture inspired by nature, designed in line with the philosophy to \\\"live with nature.\\\"\\n\\nFor weekend getaways, Shizen Energy Holidays, or department offsites, employees can stay at SANU locations situated by Japan's beautiful seas, mountains, and lakes, and enjoy a relaxing time, feeling renewed by the serene surroundings.\\n\\n*Accommodation fees will apply.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#555555\",\n            \"importance\": \"dominant\",\n            \"position\": \"Right column, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard body text. The first line 'SANU 2nd Home, shared vacation house service' is bolded and colored blue (#1E5A9E). An asterisk note regarding fees is italicized and smaller.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_593_545_962_1034.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_593_545_962_1034.jpg\"\n              }\n            ],\n            \"source\": \"Photo of kayaking near a waterfall\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"Left column, vertically centered\",\n            \"image_style\": \"Rounded corners\",\n            \"container\": null,\n            \"special_instructions\": \"A photograph showing a person kayaking in red near a waterfall cascading into a cave. The image has rounded corners.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1073_742_1464_1036.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1073_742_1464_1036.jpg\"\n              }\n            ],\n            \"source\": \"Photo of a wooden cabin in the woods\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"Right column, vertically centered\",\n            \"image_style\": \"Rounded corners\",\n            \"container\": null,\n            \"special_instructions\": \"A photograph showing a wooden A-frame cabin surrounded by trees. The image has rounded corners.\"\n          }\n        },\n        {\n          \"id\": \"illustration-17-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A speech bubble containing text about crew members staying together.\",\n          \"style\": {\n            \"illustration_style\": \"Speech bubble\",\n            \"color_scheme\": \"Light blue background (#E0F2F7) with dark blue text\",\n            \"position\": \"Bottom-right corner, overlapping the cabin image\",\n            \"special_instructions\": \"A stylized speech bubble with a tail pointing towards the cabin image. It contains the text: 'Six of us (crew members) stayed together! We cooked meals, spent time surrounded by nature away from the hustle and bustle of the city, and came back completely refreshed and re-energized.'\"\n          }\n        },\n        {\n          \"id\": \"illustration-17-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative wavy pattern along the bottom edge of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"Decorative border pattern\",\n            \"color_scheme\": \"Shades of blue and white\",\n            \"position\": \"Bottom edge, spanning full width\",\n            \"special_instructions\": \"A stylized wave pattern resembling water or ocean waves, colored in varying shades of blue, acting as a footer decoration.\"\n          }\n        },\n        {\n          \"id\": \"caption-17-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"© SHIZEN ENERGY Inc. All rights reserved.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#555555\",\n            \"position\": \"Bottom-right corner, overlaid on the wave pattern\",\n            \"text_style\": null,\n            \"special_instructions\": null\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/a148d73d-3373-4db1-a250-c4ea12e4bba1_input_0.jpg", "images/multi_image/a148d73d-3373-4db1-a250-c4ea12e4bba1_input_1.jpg"], "target_image": "images/multi_image/a148d73d-3373-4db1-a250-c4ea12e4bba1_target.jpg", "width": 2048, "height": 1152}
+{"index": 154, "id": "ca73d8be-a046-42e4-94b1-c404ec335783", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Financial Analysis\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2B579A\",\n      \"secondary\": \"#D9534F\",\n      \"text\": \"#333333\",\n      \"accent\": \"#2B579A\"\n    },\n    \"special_instructions\": \"Clean, professional corporate slide design. Uses a standard sans-serif font family (likely Arial). The layout is structured with a clear header, a bulleted text section on the left, and a photo on the right. The bottom half is split into two distinct chart areas. The color scheme relies heavily on a corporate blue (#2B579A) for headers and emphasis boxes, with data lines in blue, red, green, and grey.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Analytical, Informative\",\n    \"composition_style\": \"Asymmetric Grid\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 67,\n      \"blocks\": [\n        {\n          \"id\": \"title-67\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"A Creatively Bankrupt Culture: Viacom Tentpole Award Shows Got Slimed\",\n          \"style\": {\n            \"color\": \"#2B579A\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, dark blue text. Spans two lines.\"\n          }\n        },\n        {\n          \"id\": \"body-67-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Dramatic Drop in Popularity of the VMAs and Kids' Choice Awards\\n\\n■ MTV VMA Ratings Down 40% Y/Y this Year and 18% the Prior Year¹\\n\\n■ Some Speculate this will this be the Last Year for the Expensive Award Show²\\n\\n■ Nick's Kids' Choice Ratings Were Down 28% Y/Y³\\n\\n■ While the Grammys Have Flat Viewership in the Last 5 Years and the Oscars Have Seen a 10% Drop, the VMA's are Down 55% and Kids Choice Awards are Down 50%\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points use a square symbol. Text is left-aligned. Superscript numbers indicate footnotes.\"\n          }\n        },\n        {\n          \"id\": \"image-67-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1549_338_1848_682.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1549_338_1848_682.jpg\"\n              }\n            ],\n            \"source\": \"Photo of a person covered in green slime\",\n            \"caption\": \"A Viacom Investor\"\n          },\n          \"style\": {\n            \"position\": \"top-right\",\n            \"image_style\": \"Rounded corners\",\n            \"container\": null,\n            \"special_instructions\": \"The image has rounded corners. A caption 'A Viacom Investor' is placed directly below the image in grey text.\"\n          }\n        },\n        {\n          \"id\": \"chart-67-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_167_839_1030_1280.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_167_839_1030_1280.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing VMAS & KCA Total Viewers (2010-2015)\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-left\",\n            \"image_style\": \"Standard line chart with axes\",\n            \"container\": null,\n            \"special_instructions\": \"Chart shows two lines (Blue for VMAS, Red for Kids Choice Awards). A dark blue box with white text 'Down 53% on Avg.' is overlaid on the chart. X-axis labels are 2010-2015. Y-axis has labels 8M, 11M, 4M, 5M.\"\n          }\n        },\n        {\n          \"id\": \"chart-67-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1103_796_1886_1281.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1103_796_1886_1281.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing Oscars & Grammys Total Viewers (2010-2015)\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-right\",\n            \"image_style\": \"Standard line chart with axes\",\n            \"container\": null,\n            \"special_instructions\": \"Chart shows two lines (Green for Oscars, Grey for Grammys). A dark blue box with white text 'Down 5% on Avg.' is overlaid on the chart. X-axis labels are 2010-2015. Y-axis has labels 25M, 37M, 41M.\"\n          }\n        },\n        {\n          \"id\": \"footer-67-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Source:\\n1. http://www.billboard.com/articles/news/6685050/2015-vmas-ratings-drop-miley-cyrus\\n2. http://www.billboard.com/articles/events/vma/6678718/vmas-shutting-down-rumors-2015/\\n3. http://headlineplanet.com/home/2015/03/31/ratings-nickelodeons-kids-choice-awards-show-falls-sharply-this-year/\\n\\nSpringOwl\\nAsset Management LLC\\n\\nConfidential | For Discussion Purposes Only | 67\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Footer section. Left side contains the SpringOwl logo and source links. Right side contains confidentiality notice and page number.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/ca73d8be-a046-42e4-94b1-c404ec335783_input_0.jpg", "images/multi_image/ca73d8be-a046-42e4-94b1-c404ec335783_input_1.jpg", "images/multi_image/ca73d8be-a046-42e4-94b1-c404ec335783_input_2.jpg"], "target_image": "images/multi_image/ca73d8be-a046-42e4-94b1-c404ec335783_target.jpg", "width": 2048, "height": 1582}
+{"index": 155, "id": "009779b5-5dbf-4fb7-8051-287f8d438ba8", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Tech-Culinary Hybrid\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#FF4500\",\n      \"secondary\": \"#FF69B4\",\n      \"text\": \"#000000\",\n      \"accent\": \"#FFD700\"\n    },\n    \"special_instructions\": \"The slide features a stark white background. A vertical gradient bar runs along the far left edge, transitioning from bright orange (#FF4500) at the top to hot pink (#FF69B4) at the bottom. The right side of the slide is dominated by a large, colorful digital illustration of a female chef in a kitchen setting, surrounded by floating code snippets and data visualizations, creating a 'cyber-cooking' aesthetic.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Playful, Innovative, Professional\",\n    \"composition_style\": \"Two-column Asymmetric\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 31,\n      \"blocks\": [\n        {\n          \"id\": \"title-31\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"PEAK Threat Hunter's Cookbook\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, sans-serif font. Positioned in the upper-left quadrant of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-31-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"A structured approach to closing the gap between theory and practice in Threat Hunting.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard sans-serif font, normal weight. Left-aligned. Spaced out over three lines.\"\n          }\n        },\n        {\n          \"id\": \"body-31-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Teaching analysts to cook – with Splunk!\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center, below previous body text\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard sans-serif font, normal weight. Left-aligned. Ends with an exclamation point.\"\n          }\n        },\n        {\n          \"id\": \"illustration-31-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large, detailed digital illustration of a female chef with purple hair and glasses wearing a white chef's uniform and hat. She is typing on a laptop with a 'SURge' logo. The background of the illustration features floating code snippets, data points, and kitchen items like a bottle of juice and books, blending a culinary theme with cybersecurity data visualization.\",\n          \"style\": {\n            \"illustration_style\": \"Digital Art / Vector Illustration\",\n            \"color_scheme\": \"White, Purple, Blue, Orange, Green\",\n            \"position\": \"right-side, spanning full height\",\n            \"special_instructions\": \"The illustration acts as the primary visual anchor for the right half of the slide. It has a soft, glowing aesthetic.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/009779b5-5dbf-4fb7-8051-287f8d438ba8_input_0.jpg"], "target_image": "images/multi_image/009779b5-5dbf-4fb7-8051-287f8d438ba8_target.jpg", "width": 2048, "height": 1152}
+{"index": 156, "id": "25bcadf4-e2d2-48d7-89ff-0b907c8647c4", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#F2A900\",\n      \"secondary\": \"#333333\",\n      \"text\": \"#333333\",\n      \"accent\": \"#F2A900\"\n    },\n    \"special_instructions\": \"Clean, high-contrast corporate design. The slide features a solid white background with a distinct footer banner. The footer consists of a bright yellow/orange (#F2A900) section at the very bottom, topped by a thin grey strip. The main content area is uncluttered white space. The image on the right has a dark, tech-themed aesthetic with blue bokeh effects, contrasting with the clean slide layout.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Informative\",\n    \"composition_style\": \"Asymmetric, Two-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 197,\n      \"blocks\": [\n        {\n          \"id\": \"title-197\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Query optimization in RDS\",\n          \"style\": {\n            \"color\": \"#333333\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold sans-serif font. Positioned in the upper left quadrant, leaving significant whitespace to the right.\"\n          }\n        },\n        {\n          \"id\": \"body-197-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Use indexes to accelerate SELECT statements\\n• Use EXPLAIN plans to identify the indexes you need\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard bullet points. The second bullet is indented under the first, indicating a sub-point. Text is left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"body-197-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Avoid full table scans\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard bullet point, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"body-197-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Use ANALYZE TABLE periodically\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard bullet point, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"body-197-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Simplify WHERE clauses\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard bullet point, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"body-197-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Engine-specific optimizations\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard bullet point, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1111_265_1984_761.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1111_265_1984_761.jpg\"\n              }\n            ],\n            \"source\": \"SQL code snippet on a dark background\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right-center\",\n            \"image_style\": \"The image shows a blurred background with blue bokeh lights and a dark blue/purple gradient. The SQL code is highlighted in bright cyan and pink text against a dark background.\",\n            \"container\": null,\n            \"special_instructions\": \"The image is placed on the right side of the slide, vertically centered. It appears to be a screenshot of a code editor or terminal window.\"\n          }\n        },\n        {\n          \"id\": \"footer-1\",\n          \"type\": \"text\",\n          \"role\": \"footer\",\n          \"content\": \"sundog-education.com\\ndatacumulus.com\\n© 2023-2026 All Rights Reserved Worldwide\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-left\",\n            \"container\": \"yellow banner\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"White text located on the yellow footer banner. Small font size.\"\n          }\n        },\n        {\n          \"id\": \"footer-2\",\n          \"type\": \"text\",\n          \"role\": \"footer\",\n          \"content\": \"DataCumulus\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Logo text located on the white area of the footer, next to the yellow banner.\"\n          }\n        },\n        {\n          \"id\": \"footer-3\",\n          \"type\": \"text\",\n          \"role\": \"footer\",\n          \"content\": \"Sundog Education\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Logo text located on the white area of the footer, next to DataCumulus.\"\n          }\n        },\n        {\n          \"id\": \"sidebar-1\",\n          \"type\": \"text\",\n          \"role\": \"sidebar\",\n          \"content\": \"NOT FOR DISTRIBUTION © Stephane Maerek www.datacumulus.com\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"minor\",\n            \"position\": \"far-right edge, vertical\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Vertical text running from top to bottom along the extreme right edge of the slide.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/25bcadf4-e2d2-48d7-89ff-0b907c8647c4_input_0.jpg"], "target_image": "images/multi_image/25bcadf4-e2d2-48d7-89ff-0b907c8647c4_target.jpg", "width": 2048, "height": 1152}
+{"index": 157, "id": "30fe8b9e-4b0f-4f88-b3cc-a65071bc6d60", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Geography Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#8CC63F\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#8CC63F\"\n    },\n    \"special_instructions\": \"The slide features a clean, white background with a distinct header banner. The banner is a solid lime green (#8CC63F) with a black border. The left edge of the slide has a vertical decorative bar composed of a red stripe and a blue stripe. The map uses a specific palette of pastel colors to denote regions, overlaid on a grey topographic background.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Educational\",\n    \"composition_style\": \"Asymmetric, Header-focused\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 4,\n      \"blocks\": [\n        {\n          \"id\": \"illustration-4-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A vertical decorative bar on the far left edge of the slide, consisting of a red stripe on top of a blue stripe.\",\n          \"style\": {\n            \"illustration_style\": \"decorative border\",\n            \"color_scheme\": \"Red and Blue\",\n            \"position\": \"far-left edge, spanning full height\",\n            \"special_instructions\": \"Solid colors, no gradient.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A solid lime green rectangular banner spanning the width of the slide below the top margin.\",\n          \"style\": {\n            \"illustration_style\": \"banner\",\n            \"color_scheme\": \"Lime Green (#8CC63F)\",\n            \"position\": \"top-center, below the top margin\",\n            \"special_instructions\": \"Has a thin black border around it.\"\n          }\n        },\n        {\n          \"id\": \"title-4\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Physiographic Regions of Nepal\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"centered horizontally within the green banner\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Large font size, centered.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A circular compass rose icon in the top-left corner of the map area.\",\n          \"style\": {\n            \"illustration_style\": \"icon\",\n            \"color_scheme\": \"Black and white\",\n            \"position\": \"top-left corner of the map\",\n            \"special_instructions\": \"Standard compass design.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_68_413_1643_1298.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_68_413_1643_1298.jpg\"\n              }\n            ],\n            \"source\": \"Map of Nepal showing physiographic regions\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-left, below the green banner\",\n            \"image_style\": \"Map with colored regions and black borders\",\n            \"container\": null,\n            \"special_instructions\": \"The map includes a legend on the left and a scale bar at the bottom.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A legend box located on the left side of the map.\",\n          \"style\": {\n            \"illustration_style\": \"legend\",\n            \"color_scheme\": \"Black text with colored swatches\",\n            \"position\": \"left side of the map\",\n            \"special_instructions\": \"Lists regions: High Himalaya, High Mountain, Middle Mountain, Siwalik, Terai.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-5\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A scale bar located at the bottom center of the map.\",\n          \"style\": {\n            \"illustration_style\": \"scale bar\",\n            \"color_scheme\": \"Black lines and text\",\n            \"position\": \"bottom center of the map\",\n            \"special_instructions\": \"Marks 0 and 100km.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-6\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A data table located on the right side of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"table\",\n            \"color_scheme\": \"Black text with colored background swatches\",\n            \"position\": \"right side, vertically centered\",\n            \"special_instructions\": \"Lists Area (%) for HH, HM, MM, C, T.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-7\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A logo in the top-right corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"logo\",\n            \"color_scheme\": \"Red, Green, Blue, White\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"Circular emblem with text.\"\n          }\n        },\n        {\n          \"id\": \"body-4-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"5 physiographic regions within ~ 1000 km length (E-W) and width ~200 km (N-S)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-center of the slide\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold font weight.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-8\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Page number '4' in the bottom right corner.\",\n          \"style\": {\n            \"illustration_style\": \"text\",\n            \"color_scheme\": \"Black\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Small font size.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/30fe8b9e-4b0f-4f88-b3cc-a65071bc6d60_input_0.jpg"], "target_image": "images/multi_image/30fe8b9e-4b0f-4f88-b3cc-a65071bc6d60_target.jpg", "width": 2048, "height": 1536}
+{"index": 158, "id": "749041f7-96a3-44ff-bce8-683c39770326", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Clean Corporate Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Helvetica Neue\",\n      \"secondary\": \"Helvetica Neue\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#333333\",\n      \"secondary\": \"#333333\",\n      \"text\": \"#333333\",\n      \"accent\": \"#FFFFFF\"\n    },\n    \"special_instructions\": \"The slide utilizes a stark, minimalist aesthetic with a pure white background. The design relies heavily on negative space to separate content areas. Typography is sans-serif, clean, and legible. The layout is asymmetric, balancing two large visual elements on the left with a vertical list of text on the right.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Analytical, Clean\",\n    \"composition_style\": \"Asymmetric, Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 29,\n      \"blocks\": [\n        {\n          \"id\": \"title-29\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Content Design: Study and Analysis\",\n          \"style\": {\n            \"color\": \"#333333\",\n            \"font_family\": \"Helvetica Neue\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold text serving as the main header. The colon separates the topic from the subtitle.\"\n          }\n        },\n        {\n          \"id\": \"body-29-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Visual design and content\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#666666\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, directly below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Smaller, lighter weight text acting as a subtitle or descriptor for the main title.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_93_338_575_1347.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_93_338_575_1347.jpg\"\n              }\n            ],\n            \"source\": \"Screenshot of The Home Depot mobile app interface\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left-center\",\n            \"image_style\": \"Standard smartphone bezel frame\",\n            \"container\": null,\n            \"special_instructions\": \"A vertical smartphone mockup displaying a promotional banner for appliances.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_682_340_1168_1348.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_682_340_1168_1348.jpg\"\n              }\n            ],\n            \"source\": \"Screenshot of Etsy mobile app interface\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-left, to the right of the first phone\",\n            \"image_style\": \"Standard smartphone bezel frame\",\n            \"container\": null,\n            \"special_instructions\": \"A vertical smartphone mockup displaying the Etsy app interface with a gift card promotion.\"\n          }\n        },\n        {\n          \"id\": \"body-29-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Clear CTAs\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"right-center, top of the list\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, sans-serif text. Part of a vertical list of design principles.\"\n          }\n        },\n        {\n          \"id\": \"body-29-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Concise messaging\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"right-center, below 'Clear CTAs'\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, sans-serif text. Part of a vertical list of design principles.\"\n          }\n        },\n        {\n          \"id\": \"body-29-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Clutter free\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"right-center, below 'Concise messaging'\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, sans-serif text. Part of a vertical list of design principles.\"\n          }\n        },\n        {\n          \"id\": \"body-29-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Defined hierarchy\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"right-center, bottom of the list\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, sans-serif text. Part of a vertical list of design principles.\"\n          }\n        },\n        {\n          \"id\": \"illustration-29-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A small, stylized white smartphone icon located in the bottom right corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"Line art / Flat icon\",\n            \"color_scheme\": \"White with grey text inside\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"The icon contains the text 'Mobile Best Practices' stacked vertically inside the phone screen area. It serves as a footer or branding element.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/749041f7-96a3-44ff-bce8-683c39770326_input_0.jpg", "images/multi_image/749041f7-96a3-44ff-bce8-683c39770326_input_1.jpg"], "target_image": "images/multi_image/749041f7-96a3-44ff-bce8-683c39770326_target.jpg", "width": 2048, "height": 1536}
+{"index": 159, "id": "61512fdf-9b48-4a32-9979-b8659bdb1738", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Engineering Technical\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#D80000\",\n      \"secondary\": \"#00BFFF\",\n      \"text\": \"#000000\",\n      \"accent\": \"#FFD700\"\n    },\n    \"special_instructions\": \"The slide features a clean, high-contrast academic aesthetic typical of engineering lectures. It utilizes a white background with sharp black lines for diagrams. Key terms and process steps are highlighted in bold red text. The diagrams use flat, solid colors (bright blue and red) to distinguish heat rejection and addition phases. A decorative emblem is present in the top right corner.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Technical, Educational, Professional\",\n    \"composition_style\": \"Asymmetric, Multi-panel layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 5,\n      \"blocks\": [\n        {\n          \"id\": \"title-5\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Vapor Compression cycle\",\n          \"style\": {\n            \"color\": \"#D80000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold red text centered at the very top.\"\n          }\n        },\n        {\n          \"id\": \"body-5-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Reversed Carnot Cycle (i.e. Carnot Cycle for Refrigeration Cycle)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"container\": \"light beige rectangular panel\",\n            \"container_opacity\": \"0.9\",\n            \"special_instructions\": \"Text is inside a light beige background box with a thin black border. Located directly below the main title.\"\n          }\n        },\n        {\n          \"id\": \"body-5-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"A measure of the maximum performance to be obtained from a refrigerating machine\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text located below the beige panel. Starts with a red square bullet.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_8_420_936_1380.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_8_420_936_1380.jpg\"\n              }\n            ],\n            \"source\": \"Schematic diagram of the Reversed Carnot Cycle showing components like Compressor, Turbine, and heat exchangers.\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"middle-left\",\n            \"image_style\": \"Flat vector diagram with thick black outlines and solid color fills (Blue for heat rejection, Red for heat addition, Yellow for Turbine).\",\n            \"container\": null,\n            \"special_instructions\": \"The diagram includes labels for 'Work', 'Cool Liquid', 'Heat to high temperature source', and 'Heat from low temperature sink'. Arrows indicate flow direction.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1169_501_1834_964.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1169_501_1834_964.jpg\"\n              }\n            ],\n            \"source\": \"Temperature-Entropy (T-s) diagram of the cycle.\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"middle-right\",\n            \"image_style\": \"Line graph with a rectangular cycle shaded with diagonal lines. Axes labeled 'Temperature' and 'Entropy'.\",\n            \"container\": null,\n            \"special_instructions\": \"Includes a large arrow indicating 'Net Work' pointing to the area under the curve.\"\n          }\n        },\n        {\n          \"id\": \"body-5-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"1-2: Adiabatic compression\\n2-3: Isothermal heat rejection\\n3-4: Adiabatic expansion\\n4-1: Isothermal addition of heat or isothermal expansion\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"List of process descriptions. The process numbers (1-2, 2-3, etc.) are bolded in red text.\"\n          }\n        },\n        {\n          \"id\": \"illustration-5-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative emblem or crest located in the top right corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"Emblem\",\n            \"color_scheme\": \"Green, Gold, Red, White\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"The emblem appears to be an institutional logo, possibly for NUST (National University of Sciences and Technology) based on the text at the bottom.\"\n          }\n        },\n        {\n          \"id\": \"body-5-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Mechanical Engineering Dept. CEME NUST\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Footer text indicating the department and university.\"\n          }\n        },\n        {\n          \"id\": \"body-5-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"5\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Slide number located in the bottom right corner.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/61512fdf-9b48-4a32-9979-b8659bdb1738_input_0.jpg", "images/multi_image/61512fdf-9b48-4a32-9979-b8659bdb1738_input_1.jpg"], "target_image": "images/multi_image/61512fdf-9b48-4a32-9979-b8659bdb1738_target.jpg", "width": 2048, "height": 1536}
+{"index": 160, "id": "e81271df-706d-4cc7-bb35-d0e085086ee5", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Educational Infographic\",\n    \"background\": {\n      \"color\": \"#F0F4F8\",\n      \"gradient\": \"radial-gradient(circle at 80% 20%, #E3F2FD 0%, #F0F4F8 60%)\",\n      \"texture\": \"none\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2C3E50\",\n      \"secondary\": \"#3498DB\",\n      \"text\": \"#2C3E50\",\n      \"accent\": \"#E67E22\"\n    },\n    \"special_instructions\": \"The slide features a clean, professional educational aesthetic. The background is a very light grey-blue with a subtle radial gradient in the top right corner to highlight the university logo. The layout is split into two distinct columns: the left side uses a vertical stack of isometric layers (orange, yellow, teal) to represent service models, while the right side uses a horizontal arrangement of cloud icons (blue, green, purple) for deployment models. The typography is sans-serif, bold, and highly legible.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Educational\",\n    \"composition_style\": \"Two-column Asymmetric\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 34,\n      \"blocks\": [\n        {\n          \"id\": \"title-34\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"A Visual Guide to Cloud Computing Models\",\n          \"style\": {\n            \"color\": \"#1A252F\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, bold heading centered at the top of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-34-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"CLOUD SERVICE MODELS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, below title\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Section header, bold, uppercase, aligned left.\"\n          }\n        },\n        {\n          \"id\": \"body-34-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SaaS (Software as a Service)\\nDelivers ready-to-use software accessed via a web browser, like Gmail or Zoom.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, top section\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold label followed by descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"body-34-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"PaaS (Platform as a Service)\\nOffers a platform for developers to build and deploy apps without managing infrastructure.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, middle section\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold label followed by descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"body-34-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"IaaS (Infrastructure as a Service)\\nProvides virtual servers and storage; users manage the operating system and applications.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, bottom section\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold label followed by descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_424_301_897_1057.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_424_301_897_1057.jpg\"\n              }\n            ],\n            \"source\": \"Isometric stack diagram representing SaaS, PaaS, and IaaS layers\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-left, vertically spanning the height of the service model text\",\n            \"image_style\": \"Isometric 3D illustration with distinct colored layers (orange top, yellow middle, teal bottom).\",\n            \"container\": \"none\",\n            \"special_instructions\": \"The image is a composite graphic showing three stacked platforms. The top platform (SaaS) is orange with devices. The middle (PaaS) is yellow with code blocks and gears. The bottom (IaaS) is teal with server racks.\"\n          }\n        },\n        {\n          \"id\": \"body-34-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"CLOUD DEPLOYMENT MODELS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-right, below title\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Section header, bold, uppercase, aligned right.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_975_334_1374_572.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_975_334_1374_572.jpg\"\n              }\n            ],\n            \"source\": \"Cloud icon representing Public Cloud\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, top\",\n            \"image_style\": \"Flat vector cloud icon in light blue with user avatars and logos inside.\",\n            \"container\": \"none\",\n            \"special_instructions\": \"The cloud icon is light blue (#5DADE2) with a white interior.\"\n          }\n        },\n        {\n          \"id\": \"body-34-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Public Cloud\\nServices are owned by a third party and shared by multiple organizations.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, below Public Cloud image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold label followed by descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1439_488_1811_715.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1439_488_1811_715.jpg\"\n              }\n            ],\n            \"source\": \"Cloud icon representing Private Cloud\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, middle\",\n            \"image_style\": \"Flat vector cloud icon in light green with a large padlock in the center.\",\n            \"container\": \"none\",\n            \"special_instructions\": \"The cloud icon is light green (#88D8B0) with a white interior.\"\n          }\n        },\n        {\n          \"id\": \"body-34-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Private Cloud\\nCloud infrastructure is used exclusively by a single organization for greater control.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, below Private Cloud image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold label followed by descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"figure-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1157_723_1570_961.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1157_723_1570_961.jpg\"\n              }\n            ],\n            \"source\": \"Cloud icon representing Hybrid Cloud\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, bottom\",\n            \"image_style\": \"Flat vector cloud icon in light purple with interlocking rings and smaller cloud icons.\",\n            \"container\": \"none\",\n            \"special_instructions\": \"The cloud icon is light purple (#A29BFE) with a white interior.\"\n          }\n        },\n        {\n          \"id\": \"body-34-8\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Hybrid Cloud\\nCombines private and public clouds to balance security and cost-effectiveness.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, below Hybrid Cloud image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold label followed by descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"illustration-34-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"University logo and text in the top right corner\",\n          \"style\": {\n            \"illustration_style\": \"Logo and text block\",\n            \"color_scheme\": \"Green and Red\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"Contains the Silver Oak University crest (green tree) and text 'SILVER OAK UNIVERSITY' in red and black.\"\n          }\n        },\n        {\n          \"id\": \"illustration-34-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Small watermark text in bottom right corner\",\n          \"style\": {\n            \"illustration_style\": \"Text watermark\",\n            \"color_scheme\": \"Grey\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Small text reading 'NotebookLM'.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/e81271df-706d-4cc7-bb35-d0e085086ee5_input_0.jpg", "images/multi_image/e81271df-706d-4cc7-bb35-d0e085086ee5_input_1.jpg", "images/multi_image/e81271df-706d-4cc7-bb35-d0e085086ee5_input_2.jpg", "images/multi_image/e81271df-706d-4cc7-bb35-d0e085086ee5_input_3.jpg"], "target_image": "images/multi_image/e81271df-706d-4cc7-bb35-d0e085086ee5_target.jpg", "width": 2048, "height": 1152}
+{"index": 161, "id": "83681f12-b413-4af8-84a7-da63e13bea8e", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Tech Blue\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#1E50C2\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#1E50C2\"\n    },\n    \"special_instructions\": \"Clean, high-contrast corporate presentation style. The slide utilizes a stark white background to maximize readability. The primary brand color is a vibrant royal blue (#1E50C2) used for the main title and footer. The layout is asymmetrical, balancing a dense text column on the left with a colorful graphical process flow on the right. Decorative purple sparkles are present in the bottom right corner.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Informative, Corporate\",\n    \"composition_style\": \"Asymmetric Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 7,\n      \"blocks\": [\n        {\n          \"id\": \"title-7\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"How MaintWiz CMMS Helps Handle Shutdown Maintenance\",\n          \"style\": {\n            \"color\": \"#1E50C2\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold blue text aligned to the left margin.\"\n          }\n        },\n        {\n          \"id\": \"body-7-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Manage Shutdown Maintenance Efficiently with MaintWiz\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Sub-heading text, bold weight, black color.\"\n          }\n        },\n        {\n          \"id\": \"body-7-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"» Comprehensive asset register — track all equipment, history, criticality before shutdown\\n» Advanced shutdown planning tools — scheduling, resource allocation, risk assessment, dependencies mapping\\n» Work-order and project-management capabilities — multiple tasks handled concurrently with full visibility\\n» Real-time progress tracking, monitoring spares, manpower, and safety compliance during shutdown execution\\n» Post-shutdown analytics — performance validation, maintenance history logging, continuous improvement for next turnaround\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center, below sub-heading\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bulleted list with a double chevron (») bullet style. Standard weight text.\"\n          }\n        },\n        {\n          \"id\": \"illustration-7-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A horizontal process flow diagram consisting of four rounded square icons connected by dotted lines. The icons are colored in a gradient spectrum: Blue (Calendar icon labeled 'PLAN'), Orange (Clipboard icon labeled 'TRACK'), Purple (Wrench/Gear icon labeled 'EXECUTE'), and Green (Checklist icon labeled 'CLOSE'). Below the icons is a horizontal pill-shaped bar containing the text 'PLAN • TRACK • EXECUTE • CLOSE' in white text, matching the colors of the icons above.\",\n          \"style\": {\n            \"illustration_style\": \"Infographic element / Process flow\",\n            \"color_scheme\": \"Blue, Orange, Purple, Green gradient\",\n            \"position\": \"right-center\",\n            \"special_instructions\": \"The icons have white line art. The connecting lines are dotted. The bottom bar has a slight gradient matching the icons.\"\n          }\n        },\n        {\n          \"id\": \"caption-7-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Shutdown Maintenance — Simplified with MaintWiz\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"right-center, below the process flow graphic\",\n            \"text_style\": null,\n            \"special_instructions\": \"Centered text relative to the graphic above it.\"\n          }\n        },\n        {\n          \"id\": \"illustration-7-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Two small purple four-pointed star/sparkle shapes located in the bottom right corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"Decorative element\",\n            \"color_scheme\": \"Purple\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Small decorative accents.\"\n          }\n        },\n        {\n          \"id\": \"footer-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"© MaintWiz Technologies 2025. All rights reserved. https://www.maintwiz.com\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-left\",\n            \"container\": \"Blue arrow-shaped banner\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"White text on a solid blue background that extends to the right in an arrow shape.\"\n          }\n        },\n        {\n          \"id\": \"page-number-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"7\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right\",\n            \"container\": \"Purple star shape\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"White number inside the purple star shape.\"\n          }\n        },\n        {\n          \"id\": \"logo-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"MaintWiz\\nInnovating Industrial AI\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1E50C2\",\n            \"importance\": \"minor\",\n            \"position\": \"top-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Logo text with a purple gear icon to the left of 'MaintWiz'.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/83681f12-b413-4af8-84a7-da63e13bea8e_input_0.jpg"], "target_image": "images/multi_image/83681f12-b413-4af8-84a7-da63e13bea8e_target.jpg", "width": 2048, "height": 1152}
+{"index": 162, "id": "8d871fac-046f-43b7-bdde-b8edd062bc32", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Technical Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#333333\",\n      \"secondary\": \"#666666\",\n      \"text\": \"#333333\",\n      \"accent\": \"#FF4500\"\n    },\n    \"special_instructions\": \"Clean, minimalist corporate design. High contrast black text on white background. Uses standard sans-serif fonts. The slide relies on a diagram on the right to visualize the technical process.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Informative\",\n    \"composition_style\": \"Two-column layout with title and main text on the left, and a large diagram on the right.\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 24,\n      \"blocks\": [\n        {\n          \"id\": \"title-24\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Major version upgrade\",\n          \"style\": {\n            \"color\": \"#333333\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold font size, no background container.\"\n          }\n        },\n        {\n          \"id\": \"body-24-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Latest major version might not be available as soon as the community version\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style.\"\n          }\n        },\n        {\n          \"id\": \"body-24-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Upgrading to next major version might fail if:\\n\\n  ○ The current version is not the latest minor version\\n  ○ The current instance class is not supported (only current and next generation classes are supported)\\n  ○ Custom parameter group is not created for the next version\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below first bullet\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style with nested sub-bullets using hollow circles.\"\n          }\n        },\n        {\n          \"id\": \"body-24-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Restore a test instance from snapshot\\n\\nAttempt an upgrade on the test instance\\n\\nThe provider does pre-check for upgrade and report on issues\\n\\nVerify application compatibility and deprecated features\\n\\nUpgrade secondary nodes prior upgrading Primary\\n\\nPromote a Secondary node\\n\\nExtended support for EOL version ($$)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below nested list\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"List of bullet points.\"\n          }\n        },\n        {\n          \"id\": \"footer-24\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Pythian Services Inc. | Confidential | 24\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#999999\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small, light grey text indicating company name, confidentiality, and slide number.\"\n          }\n        },\n        {\n          \"id\": \"logo-24\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"love your data\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small logo text with a black circle icon next to it.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1394_329_1975_954.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1394_329_1975_954.jpg\"\n              }\n            ],\n            \"source\": \"MySQL Upgrade Diagram\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, vertically centered\",\n            \"image_style\": \"Diagram with blue database icons, orange load balancer icon, and arrows indicating data flow.\",\n            \"container\": null,\n            \"special_instructions\": \"Diagram shows a transition from MySQL 5.7 to 8.0 with a replica setup.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/8d871fac-046f-43b7-bdde-b8edd062bc32_input_0.jpg"], "target_image": "images/multi_image/8d871fac-046f-43b7-bdde-b8edd062bc32_target.jpg", "width": 2048, "height": 1152}
+{"index": 163, "id": "3592cd8c-557c-4501-ac6b-241efebb777a", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Indonesian Culinary Showcase\",\n    \"background\": {\n      \"color\": \"#E67E22\",\n      \"gradient\": \"linear-gradient(180deg, #E67E22 0%, #3E2723 50%, #281815 100%)\",\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#E67E22\",\n      \"secondary\": \"#3E2723\",\n      \"text\": \"#FFFFFF\",\n      \"accent\": \"#F1C40F\"\n    },\n    \"special_instructions\": \"The slide features a bold, warm color scheme dominated by burnt orange and dark brown. The background is split horizontally: the top half is a solid vibrant orange, while the bottom half transitions into a dark brown section. A subtle, horizontal striped texture is visible at the very bottom edge of the slide. The design uses rounded corners for images and text containers to create a friendly, approachable feel.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Inviting, Appetizing, Vibrant\",\n    \"composition_style\": \"Centered Title, Horizontal Grid Layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 3,\n      \"blocks\": [\n        {\n          \"id\": \"title-3\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"SIGNATURE DISHES YOU MUST TRY\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, uppercase text centered at the top of the orange background section.\"\n          }\n        },\n        {\n          \"id\": \"body-3-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Popular Indonesian dishes are a flavorful fusion of spices and fresh ingredients, offering a delicious array of options like Nasi Goreng, Rendang, and Sate Ayam that showcase the country's rich culinary traditions.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"centered below title\",\n            \"container\": \"rounded rectangle with thin white border\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is white, left-aligned within its container, which is centered horizontally on the slide.\"\n          }\n        },\n        {\n          \"id\": \"illustration-3-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Two stylized, four-pointed starburst shapes (sparkles) located to the right of the main title.\",\n          \"style\": {\n            \"illustration_style\": \"flat icon\",\n            \"color_scheme\": \"white with orange glow\",\n            \"position\": \"top-right, next to the title\",\n            \"special_instructions\": \"Decorative elements adding a touch of sparkle to the header area.\"\n          }\n        },\n        {\n          \"id\": \"visual_elements-3-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_127_558_531_850.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_127_558_531_850.jpg\"\n              }\n            ],\n            \"source\": \"Nasi Goreng (Fried Rice)\",\n            \"caption\": \"Nasi Goreng (Fried Rice)\\nA fragrant mix of rice, meat, and vegetables, seasoned with sweet soy sauce\"\n          },\n          \"style\": {\n            \"position\": \"bottom-left quadrant\",\n            \"image_style\": \"rounded corners\",\n            \"container\": \"white background\",\n            \"special_instructions\": \"Image has a white background and rounded corners. Caption text is white and positioned directly below the image.\"\n          }\n        },\n        {\n          \"id\": \"visual_elements-3-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_589_558_991_850.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_589_558_991_850.jpg\"\n              }\n            ],\n            \"source\": \"Rendang\",\n            \"caption\": \"Rendang\\nA slow-cooked beef dish in coconut milk and rich spices\"\n          },\n          \"style\": {\n            \"position\": \"bottom-center-left\",\n            \"image_style\": \"rounded corners\",\n            \"container\": \"white background\",\n            \"special_instructions\": \"Image has a white background and rounded corners. Caption text is white and positioned directly below the image.\"\n          }\n        },\n        {\n          \"id\": \"visual_elements-3-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1053_559_1459_851.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1053_559_1459_851.jpg\"\n              }\n            ],\n            \"source\": \"Sate Ayam (Chicken Skewers)\",\n            \"caption\": \"Sate Ayam (Chicken Skewers)\\nGrilled chicken served with peanut sauce\"\n          },\n          \"style\": {\n            \"position\": \"bottom-center-right\",\n            \"image_style\": \"rounded corners\",\n            \"container\": \"white background\",\n            \"special_instructions\": \"Image has a white background and rounded corners. Caption text is white and positioned directly below the image.\"\n          }\n        },\n        {\n          \"id\": \"visual_elements-3-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1518_560_1914_848.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1518_560_1914_848.jpg\"\n              }\n            ],\n            \"source\": \"Gado-Gado\",\n            \"caption\": \"Gado-Gado\\nA traditional Indonesian salad with peanut dressing\"\n          },\n          \"style\": {\n            \"position\": \"bottom-right quadrant\",\n            \"image_style\": \"rounded corners\",\n            \"container\": \"white background\",\n            \"special_instructions\": \"Image has a white background and rounded corners. Caption text is white and positioned directly below the image.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/3592cd8c-557c-4501-ac6b-241efebb777a_input_0.jpg", "images/multi_image/3592cd8c-557c-4501-ac6b-241efebb777a_input_1.jpg", "images/multi_image/3592cd8c-557c-4501-ac6b-241efebb777a_input_2.jpg", "images/multi_image/3592cd8c-557c-4501-ac6b-241efebb777a_input_3.jpg"], "target_image": "images/multi_image/3592cd8c-557c-4501-ac6b-241efebb777a_target.jpg", "width": 2048, "height": 1152}
+{"index": 164, "id": "dcfc7194-3f90-48f2-971e-3072cd391205", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Clean Data Visualization\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Helvetica Neue\",\n      \"secondary\": \"Helvetica Neue\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#000000\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#000000\"\n    },\n    \"special_instructions\": \"The slide features a stark, minimalist design with a pure white background. The typography is sans-serif, likely Helvetica or Arial, ensuring high readability. The visual weight is distributed between the bold title and the two distinct line charts. There are no background textures or gradients.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Analytical\",\n    \"composition_style\": \"Two-column asymmetric layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 57,\n      \"blocks\": [\n        {\n          \"id\": \"title-57\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"S-curve as usual, not business as usual\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Helvetica Neue\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold font size. Positioned at the very top left of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-57-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Successful technologies grow along S-curves\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, directly below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Regular weight, smaller than title but larger than body text. Acts as a subtitle.\"\n          }\n        },\n        {\n          \"id\": \"body-57-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Individual products\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below subtitle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold weight, serves as a section header for the left chart.\"\n          }\n        },\n        {\n          \"id\": \"body-57-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Technological adoption by household in the United States\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"left column, below 'Individual products'\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Regular weight, smaller font size. Acts as a chart caption.\"\n          }\n        },\n        {\n          \"id\": \"body-57-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Infrastructure systems\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, below subtitle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold weight, serves as a section header for the right chart.\"\n          }\n        },\n        {\n          \"id\": \"body-57-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Share of maximum size in the United States\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"right column, below 'Infrastructure systems'\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Regular weight, smaller font size. Acts as a chart caption.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_48_327_991_1030.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_48_327_991_1030.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing technological adoption by household (S-curves for Stove, Landline, Car, Radio, Refrigerator, Washing machine, Color TV, Cellphone, Internet)\",\n            \"caption\": \"Technological adoption by household in the United States\"\n          },\n          \"style\": {\n            \"position\": \"left column, vertically centered\",\n            \"image_style\": \"Line chart with thin black lines, grey grid lines, and multi-colored data lines (orange, red, green, blue, black).\",\n            \"container\": null,\n            \"special_instructions\": \"The chart has a white background. The X-axis ranges from 1900 to 2010. The Y-axis ranges from 0 to 100%.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1068_328_2013_1030.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1068_328_2013_1030.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing share of maximum size for infrastructure systems (S-curves for Canals, Railways, Telegraphs, Oil pipelines, Roads)\",\n            \"caption\": \"Share of maximum size in the United States\"\n          },\n          \"style\": {\n            \"position\": \"right column, vertically centered\",\n            \"image_style\": \"Line chart with thin black lines, grey grid lines, and multi-colored data lines (red, green, light blue, dark blue, yellow).\",\n            \"container\": null,\n            \"special_instructions\": \"The chart has a white background. The X-axis ranges from 1800 to 2000. The Y-axis ranges from 0 to 100%.\"\n          }\n        },\n        {\n          \"id\": \"caption-57-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Sources: Comin & Hobijn via OWID, Grubler, RMI.\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#000000\",\n            \"position\": \"bottom-left\",\n            \"text_style\": null,\n            \"special_instructions\": \"Small font size, regular weight. Located below the left chart.\"\n          }\n        },\n        {\n          \"id\": \"illustration-57-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A solid dark blue horizontal bar located at the very top left of the slide, above the main title.\",\n          \"style\": {\n            \"illustration_style\": \"solid bar\",\n            \"color_scheme\": \"dark blue (#004466)\",\n            \"position\": \"top-left corner\",\n            \"special_instructions\": \"The bar is rectangular and spans approximately 10-15% of the slide width. It serves as a visual anchor or brand element.\"\n          }\n        },\n        {\n          \"id\": \"illustration-57-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A solid light blue horizontal bar located at the top of the right column, above the 'Infrastructure systems' header.\",\n          \"style\": {\n            \"illustration_style\": \"solid bar\",\n            \"color_scheme\": \"light blue (#4488CC)\",\n            \"position\": \"top of right column\",\n            \"special_instructions\": \"The bar is rectangular and spans approximately 10-15% of the slide width. It visually separates the two columns.\"\n          }\n        },\n        {\n          \"id\": \"illustration-57-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"The EMBER logo consisting of the word 'EMBER' in bold black sans-serif font followed by a stylized green square icon.\",\n          \"style\": {\n            \"illustration_style\": \"logo\",\n            \"color_scheme\": \"black and green\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"The logo is placed flush with the bottom and right margins.\"\n          }\n        },\n        {\n          \"id\": \"illustration-57-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A thin vertical grey line separating the page number from the source text at the bottom left.\",\n          \"style\": {\n            \"illustration_style\": \"divider line\",\n            \"color_scheme\": \"light grey\",\n            \"position\": \"bottom-left, between page number and source text\",\n            \"special_instructions\": \"The line is very thin and short.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/dcfc7194-3f90-48f2-971e-3072cd391205_input_0.jpg", "images/multi_image/dcfc7194-3f90-48f2-971e-3072cd391205_input_1.jpg"], "target_image": "images/multi_image/dcfc7194-3f90-48f2-971e-3072cd391205_target.jpg", "width": 2048, "height": 1152}
+{"index": 165, "id": "e4275cf5-a871-4302-b847-03c5d09a4c30", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Clean Corporate Minimalist\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#FF6B00\",\n      \"secondary\": \"#333333\",\n      \"text\": \"#555555\",\n      \"accent\": \"#0088CC\"\n    },\n    \"special_instructions\": \"The slide features a stark white background with a solid vertical orange bar (#FF6B00) on the far left edge, acting as a brand accent. The typography is clean and sans-serif, utilizing high contrast between the bold black title and the grey body text. The layout is asymmetrical, balancing a conceptual diagram on the left with a block of text on the right.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Analytical, Informative\",\n    \"composition_style\": \"Two-column Asymmetric\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 3,\n      \"blocks\": [\n        {\n          \"id\": \"title-3\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"The messy truth\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, black text. Positioned in the upper left quadrant, aligned with the orange sidebar.\"\n          }\n        },\n        {\n          \"id\": \"subtitle-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Journeys wander and backtrack and explore\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#666666\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Smaller, grey text directly below the main title. Serves as a subtitle.\"\n          }\n        },\n        {\n          \"id\": \"illustration-3-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A conceptual diagram illustrating a non-linear customer journey. It features a large dotted circle labeled 'Exposure' containing two overlapping circles: a yellow one labeled 'Exploration' and a green one labeled 'Evaluation'. A red line starts at a point labeled 'TRIGGERS' at the top, curves down through the intersection of the circles, and ends at a small dotted circle at the bottom labeled 'PURCHASE' within a larger area labeled 'Experience'.\",\n          \"style\": {\n            \"illustration_style\": \"Conceptual Diagram\",\n            \"color_scheme\": \"Yellow, Green, Red, Blue, Black text on white\",\n            \"position\": \"center-left\",\n            \"special_instructions\": \"The diagram uses thin lines and simple geometric shapes. The 'Exposure' boundary is a large dotted circle. The 'Experience' area is a smaller dotted circle at the bottom.\"\n          }\n        },\n        {\n          \"id\": \"body-3-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"\\\"We know that what happens between trigger and purchase decision-making is not linear. We know there is a complicated web of touchpoints that differs from person to person.\\\"\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"dominant\",\n            \"position\": \"right half, vertically centered\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Block quote text. Black color. Standard paragraph spacing.\"\n          }\n        },\n        {\n          \"id\": \"body-3-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"How people decide what to buy lies in the 'messy middle' of the purchase journey\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#0088CC\",\n            \"importance\": \"supporting\",\n            \"position\": \"right half, below quote\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Hyperlinked text style. Blue color (#0088CC). Italicized.\"\n          }\n        },\n        {\n          \"id\": \"body-3-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Alistair Rennie, Jonny Protheroe/July 2020\\nThink with Google\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"minor\",\n            \"position\": \"right half, bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Attribution text. Black color. 'Think with Google' is italicized.\"\n          }\n        },\n        {\n          \"id\": \"footer-3-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"© Avalara\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#666666\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-left corner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small copyright text in grey.\"\n          }\n        },\n        {\n          \"id\": \"footer-3-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"3\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#666666\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right corner\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small slide number in grey.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/e4275cf5-a871-4302-b847-03c5d09a4c30_input_0.jpg"], "target_image": "images/multi_image/e4275cf5-a871-4302-b847-03c5d09a4c30_target.jpg", "width": 2048, "height": 1152}
+{"index": 166, "id": "54de47a7-2ce1-448c-957d-f35d950aedae", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Corporate Collaboration\",\n    \"background\": {\n      \"color\": \"#B583A6\",\n      \"gradient\": \"none\",\n      \"texture\": \"none\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Helvetica Neue\",\n      \"secondary\": \"Helvetica Neue\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#B583A6\",\n      \"secondary\": \"#E6D4F0\",\n      \"text\": \"#FDFDFD\",\n      \"accent\": \"#E6D4F0\"\n    },\n    \"special_instructions\": \"The slide features a flat, vector-based design with a dominant mauve/purple background. It utilizes a 'card' layout where the main content area is a solid block of color. Decorative elements include geometric patterns (dotted lines, crosses) in the corners and stylized 3D-style illustrations of diverse people collaborating, adding a playful yet professional tone.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Collaborative, Playful\",\n    \"composition_style\": \"Asymmetric, Content-Centered\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 16,\n      \"blocks\": [\n        {\n          \"id\": \"title-16\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"REGINA FASOLD\",\n          \"style\": {\n            \"color\": \"#FDFDFD\",\n            \"font_family\": \"Helvetica Neue\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Text is white, uppercase, sans-serif, placed inside a small, semi-transparent white rectangular box with rounded corners.\"\n          }\n        },\n        {\n          \"id\": \"body-16-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"So, help your colleagues solve different problems at the work place.\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#FDFDFD\",\n            \"importance\": \"dominant\",\n            \"position\": \"center-left\",\n            \"container\": \"solid mauve background panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Bullet point text, white color, italicized font style. Positioned in the upper-middle section of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-16-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Actively participate in different team-building efforts initiated by the organization.\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#FDFDFD\",\n            \"importance\": \"dominant\",\n            \"position\": \"center-left\",\n            \"container\": \"solid mauve background panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Bullet point text, white color, italicized font style. Positioned in the lower-middle section of the slide.\"\n          }\n        },\n        {\n          \"id\": \"illustration-16-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A stylized 3D illustration of two people brainstorming. One person sits on a stool pointing at a purple board with shapes, while another stands pointing at a grid. Thought bubbles above them contain a pretzel and a lightbulb.\",\n          \"style\": {\n            \"illustration_style\": \"3D vector art\",\n            \"color_scheme\": \"Purple, orange, blue, yellow\",\n            \"position\": \"top-right\",\n            \"special_instructions\": \"Located in the upper right corner, overlapping the main content area slightly.\"\n          }\n        },\n        {\n          \"id\": \"illustration-16-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A group of diverse 3D characters collaborating around a laptop displaying a bar chart. Elements include a large red arrow, a calculator, a notepad, and a potted plant.\",\n          \"style\": {\n            \"illustration_style\": \"3D vector art\",\n            \"color_scheme\": \"Green, red, blue, yellow, orange\",\n            \"position\": \"bottom-left\",\n            \"special_instructions\": \"Located in the bottom left corner, overlapping the main content area slightly.\"\n          }\n        },\n        {\n          \"id\": \"illustration-16-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative geometric pattern consisting of small white 'x' marks arranged in a grid.\",\n          \"style\": {\n            \"illustration_style\": \"Geometric pattern\",\n            \"color_scheme\": \"White on mauve\",\n            \"position\": \"left edge, vertically centered\",\n            \"special_instructions\": \"Acts as a vertical border element on the left side of the slide.\"\n          }\n        },\n        {\n          \"id\": \"illustration-16-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative geometric pattern consisting of small white dots arranged in a grid.\",\n          \"style\": {\n            \"illustration_style\": \"Geometric pattern\",\n            \"color_scheme\": \"White on mauve\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"Acts as a corner accent element.\"\n          }\n        },\n        {\n          \"id\": \"illustration-16-5\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative geometric pattern consisting of small white dots arranged in a grid.\",\n          \"style\": {\n            \"illustration_style\": \"Geometric pattern\",\n            \"color_scheme\": \"White on mauve\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Acts as a corner accent element.\"\n          }\n        },\n        {\n          \"id\": \"illustration-16-6\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A horizontal dotted line separating the top and bottom text sections.\",\n          \"style\": {\n            \"illustration_style\": \"Line art\",\n            \"color_scheme\": \"White\",\n            \"position\": \"middle-right\",\n            \"special_instructions\": \"Horizontal separator line.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/54de47a7-2ce1-448c-957d-f35d950aedae_input_0.jpg", "images/multi_image/54de47a7-2ce1-448c-957d-f35d950aedae_input_1.jpg"], "target_image": "images/multi_image/54de47a7-2ce1-448c-957d-f35d950aedae_target.jpg", "width": 2048, "height": 1152}
+{"index": 167, "id": "549843e2-b179-4827-a8c7-2538e4c722d8", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Corporate Clean\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#3B4A6B\",\n      \"secondary\": \"#3B4A6B\",\n      \"text\": \"#3B4A6B\",\n      \"accent\": \"#3B4A6B\"\n    },\n    \"special_instructions\": \"The slide utilizes a clean, minimalist aesthetic with a stark white background. The typography is sans-serif, likely Arial or Helvetica, ensuring high readability. The color palette is monochromatic, relying on a deep slate blue for all text elements, creating a professional and serious tone. The layout is structured with a clear visual hierarchy: a large centered title, a smaller introductory sentence, and a three-column grid of content blocks below.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Informative, Clean\",\n    \"composition_style\": \"Grid-based, Centered\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 8,\n      \"blocks\": [\n        {\n          \"id\": \"title-8\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Global Perspective: Models that Work\",\n          \"style\": {\n            \"color\": \"#3B4A6B\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large font size, centered horizontally, significant margin from the top edge.\"\n          }\n        },\n        {\n          \"id\": \"body-8-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Examples from high-performing systems demonstrate scalable practices:\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3B4A6B\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, centered\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Smaller font size than title, centered horizontally, acts as a lead-in sentence.\"\n          }\n        },\n        {\n          \"id\": \"image-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_100_366_692_733.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_100_366_692_733.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a lab setting with a researcher at a laptop and a colleague in the background\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left column of the three-column grid\",\n            \"image_style\": \"flat vector illustration style with teal and blue tones\",\n            \"container\": null,\n            \"special_instructions\": \"The image is rectangular and sits above the text block for 'Industrial PhD (Sweden)'. It has no visible border or shadow.\"\n          }\n        },\n        {\n          \"id\": \"body-8-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Industrial PhD (Sweden)\\n\\nStructured split-time programmes where doctoral candidates work on industry problems while pursuing academic rigour.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3B4A6B\",\n            \"importance\": \"supporting\",\n            \"position\": \"below the first image, left column\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The title 'Industrial PhD (Sweden)' is bolded and larger than the description paragraph below it. The text is left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"image-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_725_366_1320_733.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_725_366_1320_733.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a group of diverse people networking around a table with laptops\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center column of the three-column grid\",\n            \"image_style\": \"flat vector illustration style with teal, blue, and pink accents\",\n            \"container\": null,\n            \"special_instructions\": \"The image is rectangular and sits above the text block for 'University Spinouts (UK/US)'. It has no visible border or shadow.\"\n          }\n        },\n        {\n          \"id\": \"body-8-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"University Spinouts (UK/US)\\n\\nTargeted support for commercialisation — incubators, seed funding, and mentorship turning research into ventures.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3B4A6B\",\n            \"importance\": \"supporting\",\n            \"position\": \"below the second image, center column\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The title 'University Spinouts (UK/US)' is bolded and larger than the description paragraph below it. The text is left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"image-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1350_365_1947_733.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1350_365_1947_733.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a modern open-plan office with city skyline view and people working\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column of the three-column grid\",\n            \"image_style\": \"flat vector illustration style with blue and teal tones\",\n            \"container\": null,\n            \"special_instructions\": \"The image is rectangular and sits above the text block for 'Public-Private Research Centres (South Korea)'. It has no visible border or shadow.\"\n          }\n        },\n        {\n          \"id\": \"body-8-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Public–Private Research Centres (South Korea)\\n\\nLong-term consortia align national strategy, industry needs and academic expertise for sustained innovation.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3B4A6B\",\n            \"importance\": \"supporting\",\n            \"position\": \"below the third image, right column\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"The title 'Public–Private Research Centres (South Korea)' is bolded and larger than the description paragraph below it. The text is left-aligned.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/549843e2-b179-4827-a8c7-2538e4c722d8_input_0.jpg", "images/multi_image/549843e2-b179-4827-a8c7-2538e4c722d8_input_1.jpg", "images/multi_image/549843e2-b179-4827-a8c7-2538e4c722d8_input_2.jpg"], "target_image": "images/multi_image/549843e2-b179-4827-a8c7-2538e4c722d8_target.jpg", "width": 2048, "height": 1152}
+{"index": 168, "id": "840abd2f-4d52-4ec1-aa93-a0792fa91273", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Professional Contact Slide\",\n    \"background\": {\n      \"color\": \"#1A2342\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#0078D7\",\n      \"secondary\": \"#4BC0C0\",\n      \"text\": \"#FFFFFF\",\n      \"accent\": \"#1A2342\"\n    },\n    \"special_instructions\": \"The slide features a split-screen layout. The left side is a full-height vertical strip containing a dark blue background with a cyan accent bar and the word 'THANKS' in large, semi-transparent blue text. The right side is a dark blue background hosting a large, rounded-corner blue content card. The overall aesthetic is clean, corporate, and modern.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Corporate, Clean\",\n    \"composition_style\": \"Split-screen (Left decorative/branding, Right content card)\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 13,\n      \"blocks\": [\n        {\n          \"id\": \"title-13\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Get in Touch!\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center within the right-side blue card\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, white, sans-serif font. Acts as the main heading for the contact information.\"\n          }\n        },\n        {\n          \"id\": \"body-13-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"VBEYOND CORPORATION HEADQUARTERS\\nAmwell Commons, 390 Amwell Road, Bldg #1, Suite #107\\nHillsborough, New Jersey USA – 088 44\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"centered below the title within the blue card\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Address block. 'VBEYOND CORPORATION HEADQUARTERS' is in bold/uppercase. The address lines are in a lighter weight.\"\n          }\n        },\n        {\n          \"id\": \"body-13-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"EMAIL\\ninfo@vantedgesearch.com\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"centered below address block\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Label 'EMAIL' is bold. The email address is a standard weight, likely underlined or appearing as a hyperlink.\"\n          }\n        },\n        {\n          \"id\": \"body-13-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"WEBSITE\\nwww.vantedgesearch.com\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"centered below email block\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Label 'WEBSITE' is bold. The URL is standard weight.\"\n          }\n        },\n        {\n          \"id\": \"body-13-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"PHONE\\n+1 (908) 854-0425\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"centered below website block\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Label 'PHONE' is bold. The number is standard weight.\"\n          }\n        },\n        {\n          \"id\": \"illustration-13-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large, rounded-corner rectangular card with a solid blue background (#0078D7) containing the contact information.\",\n          \"style\": {\n            \"illustration_style\": \"solid color panel\",\n            \"color_scheme\": \"Blue (#0078D7)\",\n            \"position\": \"right side of the slide, vertically centered\",\n            \"special_instructions\": \"The card has rounded corners (approx 20-30px radius). It overlaps the dark blue background.\"\n          }\n        },\n        {\n          \"id\": \"illustration-13-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A vertical decorative bar on the far left edge of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"geometric shape\",\n            \"color_scheme\": \"Dark Blue (#1A2342) with a Cyan (#4BC0C0) horizontal stripe near the top\",\n            \"position\": \"far left edge, full height\",\n            \"special_instructions\": \"The bar contains the word 'THANKS' written vertically in large, semi-transparent dark blue letters.\"\n          }\n        },\n        {\n          \"id\": \"illustration-13-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A logo in the top right corner of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"text logo with icon\",\n            \"color_scheme\": \"White and Cyan\",\n            \"position\": \"top right corner\",\n            \"special_instructions\": \"Text reads 'Vantage Search' with a stylized 'V' icon.\"\n          }\n        },\n        {\n          \"id\": \"illustration-13-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Three small white geometric shapes (triangles/arrows) at the bottom center of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"minimalist icons\",\n            \"color_scheme\": \"White\",\n            \"position\": \"bottom center, below the blue contact card\",\n            \"special_instructions\": \"Shapes appear to be downward-pointing triangles or chevrons.\"\n          }\n        },\n        {\n          \"id\": \"image-13-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_6_1_947_1150.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_6_1_947_1150.jpg\"\n              }\n            ],\n            \"source\": \"Group of business men in suits sitting on a bench\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left side of the slide, behind the vertical decorative bar\",\n            \"image_style\": \"full height, cropped\",\n            \"container\": \"none\",\n            \"special_instructions\": \"The image shows four men in business attire. It is cropped to show the left side of the scene. The vertical decorative bar overlays the left edge of this image.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/840abd2f-4d52-4ec1-aa93-a0792fa91273_input_0.jpg"], "target_image": "images/multi_image/840abd2f-4d52-4ec1-aa93-a0792fa91273_target.jpg", "width": 2048, "height": 1152}
+{"index": 169, "id": "32b16a36-9066-43de-8416-19c3beae876b", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Minimalist Corporate with Cartoon Illustration\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#000000\",\n      \"secondary\": \"#808080\",\n      \"text\": \"#000000\",\n      \"accent\": \"#D8BFD8\"\n    },\n    \"special_instructions\": \"The slide utilizes a stark white background to maximize contrast. The typography is strictly sans-serif, relying on weight (bold vs regular) and color (black vs grey) to establish hierarchy rather than distinct font families. The visual style is clean and direct, typical of modern business presentations, but juxtaposed with a colorful, high-contrast cartoon illustration.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional yet Playful\",\n    \"composition_style\": \"Asymmetric Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 6,\n      \"blocks\": [\n        {\n          \"id\": \"title-6\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"90% of CEOs do not respond to cold calls or emails.\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left quadrant\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"The title is the most prominent text element, split across two lines. It is left-aligned and uses a very heavy font weight to grab attention immediately.\"\n          }\n        },\n        {\n          \"id\": \"body-6-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Connect with customers on Twitter, in LinkedIn Groups and on blogs BEFORE you cold call them; then it won’t be so cold!\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#808080\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below the title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"This body text is left-aligned. The word 'BEFORE' is capitalized for emphasis. The color is a medium grey, significantly lighter than the title, creating a clear visual distinction between the problem (title) and the solution (body).\"\n          }\n        },\n        {\n          \"id\": \"body-6-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Source: Harvard Business Review\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"left column, below the main body text\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small, black text serving as a citation. It is left-aligned with the text block above it.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1162_369_1993_1216.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1162_369_1993_1216.jpg\"\n              }\n            ],\n            \"source\": \"Cartoon illustration of an angry man on the phone\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, vertically centered\",\n            \"image_style\": \"Flat vector art style (The Simpsons aesthetic)\",\n            \"container\": null,\n            \"special_instructions\": \"A large, colorful illustration of a man with grey hair and a bowtie shouting into a purple telephone. The image is placed on the right side of the slide, balancing the text on the left. The character's expression conveys frustration, reinforcing the text about 'cold calls'.\"\n          }\n        },\n        {\n          \"id\": \"illustration-6-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A small logo consisting of a lightbulb inside a UFO shape with the text 'MARKETINGTHINK.COM' next to it.\",\n          \"style\": {\n            \"illustration_style\": \"Line art logo\",\n            \"color_scheme\": \"Black and white\",\n            \"position\": \"bottom-left corner\",\n            \"special_instructions\": \"A watermark-style logo placed in the footer area.\"\n          }\n        },\n        {\n          \"id\": \"illustration-6-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A social media handle text '@GerryMoran' in a simple sans-serif font.\",\n          \"style\": {\n            \"illustration_style\": \"Text graphic\",\n            \"color_scheme\": \"Black\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Attribution text placed in the footer area.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/32b16a36-9066-43de-8416-19c3beae876b_input_0.jpg"], "target_image": "images/multi_image/32b16a36-9066-43de-8416-19c3beae876b_target.jpg", "width": 2048, "height": 1536}
+{"index": 170, "id": "df10633f-7318-4f42-823c-7023ca743dbd", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Academic Presentation with Sketch Infographic\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Times New Roman\",\n      \"secondary\": \"Times New Roman\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#1E3A8A\",\n      \"secondary\": \"#D4A373\",\n      \"text\": \"#1F2937\",\n      \"accent\": \"#1E3A8A\"\n    },\n    \"special_instructions\": \"The slide features a realistic wooden background texture visible at the very edges. The main content is contained within a white paper-like sheet with a thin brown border, held by two dark grey binder clips at the top and bottom. The infographic on the right uses a distinct 'blue chalk/sketch' style with heavy cross-hatching.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Academic, Informative, Creative\",\n    \"composition_style\": \"Two-column layout with a centered title\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 15,\n      \"blocks\": [\n        {\n          \"id\": \"title-15\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Public Relations Objectives for Retailers\",\n          \"style\": {\n            \"color\": \"#1F2937\",\n            \"font_family\": \"Times New Roman\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large serif font, centered horizontally.\"\n          }\n        },\n        {\n          \"id\": \"body-15-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Increase awareness of the retailer and its strategy mix\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-column, top\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style (solid circle), standard serif font.\"\n          }\n        },\n        {\n          \"id\": \"body-15-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Maintain or improve the company image\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-column, upper-middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style (solid circle).\"\n          }\n        },\n        {\n          \"id\": \"body-15-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Show the retailer as a contributor to the public's quality of life\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-column, middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style (solid circle).\"\n          }\n        },\n        {\n          \"id\": \"body-15-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Demonstrate innovativeness\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-column, middle-lower\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style (solid circle).\"\n          }\n        },\n        {\n          \"id\": \"body-15-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Present a favorable message in a highly believable manner\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-column, lower-middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style (solid circle).\"\n          }\n        },\n        {\n          \"id\": \"body-15-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Minimize total promotion costs\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#1F2937\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-column, bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style (solid circle).\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1090_419_1917_1044.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1090_419_1917_1044.jpg\"\n              }\n            ],\n            \"source\": \"Public Relations Infographic\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right-column, vertically centered\",\n            \"image_style\": \"Sketch style with blue ink, heavy cross-hatching texture on text and icons.\",\n            \"container\": null,\n            \"special_instructions\": \"The image contains the text 'PUBLIC RELATIONS' in large blue sketch font, surrounded by icons for communication, media, and events.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/df10633f-7318-4f42-823c-7023ca743dbd_input_0.jpg"], "target_image": "images/multi_image/df10633f-7318-4f42-823c-7023ca743dbd_target.jpg", "width": 2048, "height": 1152}
+{"index": 171, "id": "9bf4bff6-5a48-4932-8b3e-76ab8d83ba30", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Data Visualization\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#E67E22\",\n      \"secondary\": \"#7F8C8D\",\n      \"text\": \"#2C3E50\",\n      \"accent\": \"#3498DB\"\n    },\n    \"special_instructions\": \"Clean, minimalist corporate design. High contrast between text and background. Uses a specific shade of orange-red for section headers and a dark slate blue for main content. The visual style is flat and modern.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Informative\",\n    \"composition_style\": \"Asymmetric, Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 13,\n      \"blocks\": [\n        {\n          \"id\": \"title-13\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"SOCIAL SELLING\",\n          \"style\": {\n            \"color\": \"#E67E22\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left, below the large number '7'\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Small, uppercase text acting as a section header.\"\n          }\n        },\n        {\n          \"id\": \"body-13-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Globally, there are over 1.5 billion social media users\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"dominant\",\n            \"position\": \"left-center, below the section header\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Large, bold text for the statistic. '1.5 billion' is emphasized with bold weight.\"\n          }\n        },\n        {\n          \"id\": \"body-13-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"(McKinsey and Company)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#95A5A6\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-center, below the main statistic\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Italicized, lighter grey text indicating the source.\"\n          }\n        },\n        {\n          \"id\": \"illustration-13-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A circular graphic representing a global network. It features a stylized globe with blue oceans and white continents. Overlaid on the globe is a network graph consisting of dark blue nodes of varying sizes connected by thin black lines, symbolizing social connections.\",\n          \"style\": {\n            \"illustration_style\": \"flat icon\",\n            \"color_scheme\": \"blue, white, dark slate blue, grey\",\n            \"position\": \"right side, vertically centered\",\n            \"special_instructions\": \"The graphic is enclosed within a thick grey circular border.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/9bf4bff6-5a48-4932-8b3e-76ab8d83ba30_input_0.jpg"], "target_image": "images/multi_image/9bf4bff6-5a48-4932-8b3e-76ab8d83ba30_target.jpg", "width": 2048, "height": 1536}
+{"index": 172, "id": "89940f3a-0f51-4317-bf73-296338892728", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Minimalist Corporate\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Helvetica Neue\",\n      \"secondary\": \"Helvetica Neue\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#4ECDC4\",\n      \"secondary\": \"#F7F7F7\",\n      \"text\": \"#666666\",\n      \"accent\": \"#FFFFFF\"\n    },\n    \"special_instructions\": \"The slide features a clean, high-key aesthetic with a stark white background. The design relies on negative space and geometric shapes. The typography is sans-serif and crisp. A subtle decorative element of a 4x4 grid of light grey dots is present at the bottom left. The color palette is strictly limited to white, light grey, and a specific teal accent color.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Clean, Minimalist\",\n    \"composition_style\": \"Asymmetric Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 2,\n      \"blocks\": [\n        {\n          \"id\": \"illustration-2-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A circular logo composed of 12 colored segments arranged in a ring, located in the top-left corner.\",\n          \"style\": {\n            \"illustration_style\": \"Icon / Logo\",\n            \"color_scheme\": \"Multi-colored segments (purple, blue, teal, green, orange, red) on a white background\",\n            \"position\": \"Top-left corner\",\n            \"special_instructions\": \"Small size, acts as a brand mark.\"\n          }\n        },\n        {\n          \"id\": \"title-2\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"MEET TYPE 1: THE PERFECTIONIST\",\n          \"style\": {\n            \"color\": \"#4ECDC4\",\n            \"font_family\": \"Helvetica Neue\",\n            \"importance\": \"dominant\",\n            \"position\": \"Top-left, below the logo\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"The text is left-aligned. It is split into two lines. The color is a distinct teal/turquoise.\"\n          }\n        },\n        {\n          \"id\": \"body-2-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Strives for perfection and moral integrity\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#666666\",\n            \"importance\": \"supporting\",\n            \"position\": \"Left column, below the title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style (solid circle). Text is left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"body-2-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Works tirelessly to improve themselves and the world around them\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#666666\",\n            \"importance\": \"supporting\",\n            \"position\": \"Left column, below the first bullet\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style (solid circle). Text is left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"body-2-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Acts as a moral compass in society\",\n          \"style\": {\n            \"font_family\": \"Helvetica Neue\",\n            \"color\": \"#666666\",\n            \"importance\": \"supporting\",\n            \"position\": \"Left column, below the second bullet\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point style (solid circle). Text is left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"illustration-2-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative pattern of 16 small, light grey dots arranged in a 4x4 grid.\",\n          \"style\": {\n            \"illustration_style\": \"Dotted Pattern\",\n            \"color_scheme\": \"Light grey (#CCCCCC) on white\",\n            \"position\": \"Bottom-left corner\",\n            \"special_instructions\": \"Subtle decorative element, low opacity.\"\n          }\n        },\n        {\n          \"id\": \"image-2-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1169_114_2047_1033.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1169_114_2047_1033.jpg\"\n              }\n            ],\n            \"source\": \"Abstract architectural photography\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"Right column, spanning full height\",\n            \"image_style\": \"Full bleed within the column area\",\n            \"container\": null,\n            \"special_instructions\": \"The image features a minimalist interior design with a white circular object, vertical dark slats, and a white ledge against a grey wall.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/89940f3a-0f51-4317-bf73-296338892728_input_0.jpg"], "target_image": "images/multi_image/89940f3a-0f51-4317-bf73-296338892728_target.jpg", "width": 2048, "height": 1152}
+{"index": 173, "id": "f8cfb475-0bff-4fd4-bf94-209f83dda23a", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Vintage Industrial Steampunk\",\n    \"background\": {\n      \"color\": \"#F0EBD9\",\n      \"gradient\": null,\n      \"texture\": \"paper\",\n      \"pattern\": \"faint technical blueprints and compass roses\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Garamond\",\n      \"secondary\": \"Garamond\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#8B7355\",\n      \"secondary\": \"#D4C5A9\",\n      \"text\": \"#2C2C2C\",\n      \"accent\": \"#8B7355\"\n    },\n    \"special_instructions\": \"The slide utilizes a sepia-toned, vintage aesthetic reminiscent of old engravings or technical schematics. The background features a subtle, light blue blueprint pattern with geometric shapes and compass roses. Text is rendered in a classic serif font (likely Garamond or similar) in dark grey/black. Key elements are highlighted with beige/cream banners that have a slight drop shadow to create depth. The imagery is monochromatic, using shades of grey, black, and white, with a hand-drawn, etched style.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Educational, Serious, Historical\",\n    \"composition_style\": \"Two-column asymmetric layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 11,\n      \"blocks\": [\n        {\n          \"id\": \"title-11\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Schools Are Factories for Forging Nations.\",\n          \"style\": {\n            \"color\": \"#2C2C2C\",\n            \"font_family\": \"Garamond\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold serif font. The text is split into two lines. The first line is slightly larger than the second.\"\n          }\n        },\n        {\n          \"id\": \"body-11-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"If not education, what is the purpose? Brainwashing. The first societies to introduce free, mandatory schooling were the most militaristic in history.\",\n          \"style\": {\n            \"font_family\": \"Garamond\",\n            \"color\": \"#2C2C2C\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard serif font, normal weight. The word 'Brainwashing' is slightly emphasized by context.\"\n          }\n        },\n        {\n          \"id\": \"body-11-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"The Historical Precedent\",\n          \"style\": {\n            \"font_family\": \"Garamond\",\n            \"color\": \"#2C2C2C\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, middle\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold serif font, acting as a sub-heading.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_109_629_977_951.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_109_629_977_951.jpg\"\n              }\n            ],\n            \"source\": \"Historical symbols representing militaristic societies\",\n            \"caption\": \"Commonality: All three were societies fundamentally dedicated to war. They needed schools to prepare citizens for obedience and sacrifice.\"\n          },\n          \"style\": {\n            \"position\": \"left column, below 'The Historical Precedent'\",\n            \"image_style\": \"Monochromatic engraving style\",\n            \"container\": \"None, but sits above a beige banner\",\n            \"special_instructions\": \"The image contains three distinct elements: a Roman-style helmet with a plume, a stepped pyramid (Chichen Itza), and a double-headed eagle emblem (Habsburg). Below the image is a beige ribbon/banner containing the caption text.\"\n          }\n        },\n        {\n          \"id\": \"body-11-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Commonality: All three were societies fundamentally dedicated to war. They needed schools to prepare citizens for obedience and sacrifice.\",\n          \"style\": {\n            \"font_family\": \"Garamond\",\n            \"color\": \"#2C2C2C\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, bottom (inside banner)\",\n            \"container\": \"Beige ribbon/banner shape\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Text is centered within a decorative beige banner that spans the width of the image above it.\"\n          }\n        },\n        {\n          \"id\": \"title-11-right\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"The Mechanism of Brainwashing\",\n          \"style\": {\n            \"color\": \"#2C2C2C\",\n            \"font_family\": \"Garamond\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-right\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold serif font, aligned to the right side of the slide.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1064_120_2011_963.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1064_120_2011_963.jpg\"\n              }\n            ],\n            \"source\": \"Diagram illustrating the three stages of brainwashing\",\n            \"caption\": \"3. INDOCTRINATION: You learn a shared language, history, and geography to believe you are the same as millions of strangers. This makes you willing to die for an idea.\"\n          },\n          \"style\": {\n            \"position\": \"right column, spanning from top to bottom\",\n            \"image_style\": \"Monochromatic engraving style with mechanical gears connecting the stages\",\n            \"container\": \"None\",\n            \"special_instructions\": \"The image is a complex infographic divided into three sections labeled '1. SEPARATION', '2. INSECURITY', and '3. INDOCTRINATION'. The sections are connected by gears and chains, symbolizing a machine. The bottom section includes the text '3. INDOCTRINATION' and the explanatory paragraph.\"\n          }\n        },\n        {\n          \"id\": \"body-11-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"You learn a shared language, history, and geography to believe you are the same as millions of strangers. This makes you willing to die for an idea.\",\n          \"style\": {\n            \"font_family\": \"Garamond\",\n            \"color\": \"#2C2C2C\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, bottom (inside the infographic)\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Text is located at the bottom of the right column, directly below the '3. INDOCTRINATION' label within the infographic.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/f8cfb475-0bff-4fd4-bf94-209f83dda23a_input_0.jpg", "images/multi_image/f8cfb475-0bff-4fd4-bf94-209f83dda23a_input_1.jpg"], "target_image": "images/multi_image/f8cfb475-0bff-4fd4-bf94-209f83dda23a_target.jpg", "width": 2048, "height": 1143}
+{"index": 174, "id": "5ccb3e33-0e7c-4a39-af44-d14510d40f72", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Data Visualization\",\n    \"background\": {\n      \"color\": \"#F2F2F2\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#1A2B4A\",\n      \"secondary\": \"#8B3A3A\",\n      \"text\": \"#1A2B4A\",\n      \"accent\": \"#8B3A3A\"\n    },\n    \"special_instructions\": \"The slide features a clean, professional corporate aesthetic. The background is a light grey (#F2F2F2) with a dark blue header bar. The chart uses a specific palette of muted, desaturated colors for the stacked bars to ensure readability and a non-jarring visual experience. The chart is framed by a thin dark blue border.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Analytical\",\n    \"composition_style\": \"Asymmetric, Header-focused\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 6,\n      \"blocks\": [\n        {\n          \"id\": \"title-6\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Fans - Geography\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Located inside a dark blue header bar spanning the top of the slide.\"\n          }\n        },\n        {\n          \"id\": \"chart-6\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_113_217_1882_1049.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_113_217_1882_1049.jpg\"\n              }\n            ],\n            \"source\": \"Stacked bar chart showing fan distribution by geography for various dating apps\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center, spanning most of the slide width\",\n            \"image_style\": \"embedded chart with gridlines and legend\",\n            \"container\": \"none\",\n            \"special_instructions\": \"The chart displays 14 dating apps on the x-axis and percentage distribution on the y-axis. The legend at the top lists countries and regions. The bars are stacked.\"\n          }\n        },\n        {\n          \"id\": \"body-6-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Dating Apps\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A2B4A\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small text label in the bottom left corner, likely indicating the category of the chart.\"\n          }\n        },\n        {\n          \"id\": \"logo-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"unmetric\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A2B4A\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Brand logo text in the bottom right corner.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/5ccb3e33-0e7c-4a39-af44-d14510d40f72_input_0.jpg"], "target_image": "images/multi_image/5ccb3e33-0e7c-4a39-af44-d14510d40f72_target.jpg", "width": 2048, "height": 1152}
+{"index": 175, "id": "0606c2e0-ed6d-4df2-858d-9aa431ac2b5f", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Data Presentation\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#0078D7\",\n      \"secondary\": \"#D9531F\",\n      \"text\": \"#000000\",\n      \"accent\": \"#1F4E79\"\n    },\n    \"special_instructions\": \"The slide features a clean, professional aesthetic typical of technical presentations. The background is stark white. A solid blue banner (#0078D7) spans the top width, containing the main title in white text. The charts utilize a standard color palette: a burnt orange (#D9531F) for 'Scalding' and a medium blue (#1F4E79) for 'Spark'. Grid lines in the charts are light grey.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Data-Driven\",\n    \"composition_style\": \"Asymmetric Grid\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 17,\n      \"blocks\": [\n        {\n          \"id\": \"title-17\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Scalding vs Spark: Daily Active User Count\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center, inside a blue banner\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Text is white, bold, and centered within a solid blue rectangular banner that spans the full width of the slide.\"\n          }\n        },\n        {\n          \"id\": \"chart-17-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_49_279_984_736.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_49_279_984_736.jpg\"\n              }\n            ],\n            \"source\": \"Bar chart comparing Task Times\",\n            \"caption\": \"Scalding vs Spark - Task Times\"\n          },\n          \"style\": {\n            \"position\": \"left-center\",\n            \"image_style\": \"Standard bar chart with grid lines\",\n            \"container\": null,\n            \"special_instructions\": \"The chart title 'Scalding vs Spark - Task Times' is positioned directly above the chart area. The legend is located in the top-right corner of the chart area.\"\n          }\n        },\n        {\n          \"id\": \"chart-17-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1127_282_1999_748.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1127_282_1999_748.jpg\"\n              }\n            ],\n            \"source\": \"Bar chart comparing Wallclock Times\",\n            \"caption\": \"Scalding vs Spark - Wallclock Times\"\n          },\n          \"style\": {\n            \"position\": \"right-center\",\n            \"image_style\": \"Standard bar chart with grid lines\",\n            \"container\": null,\n            \"special_instructions\": \"The chart title 'Scalding vs Spark - Wallclock Times' is positioned directly above the chart area. The legend is located in the top-right corner of the chart area.\"\n          }\n        },\n        {\n          \"id\": \"body-17-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"No warm-up, reading from disk\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text, black color, standard weight.\"\n          }\n        },\n        {\n          \"id\": \"body-17-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Data size: 70GB (small), 477GB (large)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text, black color, standard weight.\"\n          }\n        },\n        {\n          \"id\": \"body-17-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"2-3x speedup in task and wall-clock times\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet point text, black color, standard weight.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 2 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/0606c2e0-ed6d-4df2-858d-9aa431ac2b5f_input_0.jpg", "images/multi_image/0606c2e0-ed6d-4df2-858d-9aa431ac2b5f_input_1.jpg"], "target_image": "images/multi_image/0606c2e0-ed6d-4df2-858d-9aa431ac2b5f_target.jpg", "width": 2048, "height": 1152}
+{"index": 176, "id": "c471e7b5-e4dc-4397-8150-5c8f910ad2b6", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Dark Cinematic Sports\",\n    \"background\": {\n      \"color\": \"#1A1A1A\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#FFFFFF\",\n      \"secondary\": \"#FFFFFF\",\n      \"text\": \"#E0E0E0\",\n      \"accent\": \"#FFFFFF\"\n    },\n    \"special_instructions\": \"The slide features a dark, moody aesthetic typical of sports documentaries. The background is a solid, deep charcoal grey. The design relies heavily on high-contrast imagery within circular frames. The text is stark white, ensuring high readability against the dark background. There are no gradients or complex textures; the visual interest comes entirely from the photography and the circular masking effect applied to the portraits.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Dramatic, Iconic\",\n    \"composition_style\": \"Grid-based, Centered\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 4,\n      \"blocks\": [\n        {\n          \"id\": \"title-4\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Legendary Players Who Turned Pro\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold, sans-serif font. Slight tracking (letter-spacing) to spread the letters out for a cinematic feel.\"\n          }\n        },\n        {\n          \"id\": \"body-4-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"From College Stars to National Icons\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#E0E0E0\",\n            \"importance\": \"supporting\",\n            \"position\": \"below title, top-center\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Smaller, lighter weight sans-serif font, acting as a subtitle.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_108_302_501_704.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_108_302_501_704.jpg\"\n              }\n            ],\n            \"source\": \"Portrait of Jim Brown\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left column, vertically centered\",\n            \"image_style\": \"Circular crop (circle mask) applied to the rectangular image source\",\n            \"container\": null,\n            \"special_instructions\": \"The image is cropped into a perfect circle. The image itself is black and white (monochrome).\"\n          }\n        },\n        {\n          \"id\": \"body-4-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Jim Brown\\n\\nSyracuse University:\\nRevolutionized the NFL running back position after dominating college football in the 1950s with unprecedented athleticism.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#E0E0E0\",\n            \"importance\": \"supporting\",\n            \"position\": \"below the first image, left column\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Name is bold. University line is bold. Body text is regular weight. Left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_581_383_956_705.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_581_383_956_705.jpg\"\n              }\n            ],\n            \"source\": \"Action shot of Bill Russell dribbling\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"second column from left, vertically centered\",\n            \"image_style\": \"Circular crop (circle mask) applied to the rectangular image source\",\n            \"container\": null,\n            \"special_instructions\": \"The image is cropped into a perfect circle. The image is black and white.\"\n          }\n        },\n        {\n          \"id\": \"body-4-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Bill Russell\\n\\nUniversity of San Francisco:\\n\\nLed college basketball to\\n\\nnational prominence with two\\n\\nNCAA championships before\\n\\nhis legendary NBA Hall of Fame\\ncareer.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#E0E0E0\",\n            \"importance\": \"supporting\",\n            \"position\": \"below the second image, second column\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Name is bold. University line is bold. Body text is regular weight. Left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1037_308_1434_704.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1037_308_1434_704.jpg\"\n              }\n            ],\n            \"source\": \"Portrait of Jackie Robinson in UCLA jacket\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"third column from left, vertically centered\",\n            \"image_style\": \"Circular crop (circle mask) applied to the rectangular image source\",\n            \"container\": null,\n            \"special_instructions\": \"The image is cropped into a perfect circle. The image is in color (blue jacket).\"\n          }\n        },\n        {\n          \"id\": \"body-4-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Jackie Robinson\\n\\nUCLA: Excelled in multiple college sports before breaking baseball's color barrier and becoming an American civil rights icon.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#E0E0E0\",\n            \"importance\": \"supporting\",\n            \"position\": \"below the third image, third column\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Name is bold. University line is bold. Body text is regular weight. Left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"figure-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1515_305_1889_702.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1515_305_1889_702.jpg\"\n              }\n            ],\n            \"source\": \"Portrait of a modern football player with stadium overlay\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"fourth column (right), vertically centered\",\n            \"image_style\": \"Circular crop (circle mask) applied to the rectangular image source\",\n            \"container\": null,\n            \"special_instructions\": \"The image is cropped into a perfect circle. The image is in color with a digital stadium overlay effect on the jersey.\"\n          }\n        },\n        {\n          \"id\": \"body-4-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Modern Legends\\n\\nRecent Stars: Players like Tim Tebow (Florida) and Breanna Stewart (UConn) became household names through college success before professional stardom.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#E0E0E0\",\n            \"importance\": \"supporting\",\n            \"position\": \"below the fourth image, right column\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Name is bold. Sub-header 'Recent Stars:' is bold. Body text is regular weight. Left-aligned.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/c471e7b5-e4dc-4397-8150-5c8f910ad2b6_input_0.jpg", "images/multi_image/c471e7b5-e4dc-4397-8150-5c8f910ad2b6_input_1.jpg", "images/multi_image/c471e7b5-e4dc-4397-8150-5c8f910ad2b6_input_2.jpg", "images/multi_image/c471e7b5-e4dc-4397-8150-5c8f910ad2b6_input_3.jpg"], "target_image": "images/multi_image/c471e7b5-e4dc-4397-8150-5c8f910ad2b6_target.jpg", "width": 2048, "height": 1154}
+{"index": 177, "id": "d07410e8-bc20-4912-86d1-b92c47157eb3", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Travel & Tourism\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Times New Roman\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#006080\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#006080\"\n    },\n    \"special_instructions\": \"The slide features a clean, high-contrast design with a stark white background. The visual style is defined by a deep teal accent color (#006080) used for the title and decorative sidebars. The layout is split asymmetrically: text-heavy on the left, image-heavy on the right. The typography mixes a bold, modern sans-serif for the header with a classic serif for the body text, creating a sophisticated yet accessible look. The images are framed by thin black lines and teal bars, creating a collage-like effect.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Educational, Travel-oriented\",\n    \"composition_style\": \"Asymmetric Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 4,\n      \"blocks\": [\n        {\n          \"id\": \"title-4\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"SITE\",\n          \"style\": {\n            \"color\": \"#006080\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left, inside a black rectangular border\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": \"panel\",\n            \"special_instructions\": \"The title is uppercase, widely spaced (kerning), and enclosed in a thin black rectangular border that extends to the right edge of the text area.\"\n          }\n        },\n        {\n          \"id\": \"body-4-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"is the word used for interesting places to see, often visited by tourists. Sites can be any place where something happened, or that is used for something (e.g. an archeological site, the site of the battle, a camping site). The phrase that is normally used is \\\"sights to see.\\\"\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below the title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Standard paragraph text. The word 'Sites' is bolded for emphasis. The text is left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Two vertical teal rectangular bars acting as sidebars or dividers. One is located on the far left edge of the image section, and the other is on the far right edge. They appear to be part of the frame design.\",\n          \"style\": {\n            \"illustration_style\": \"decorative shape\",\n            \"color_scheme\": \"solid teal (#006080)\",\n            \"position\": \"left and right edges of the image section, vertically centered\",\n            \"special_instructions\": \"These bars serve as visual anchors for the image layout.\"\n          }\n        },\n        {\n          \"id\": \"illustration-4-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A thin black rectangular border framing the entire image section on the right side of the slide. It has an open top and bottom, enclosing the two photos.\",\n          \"style\": {\n            \"illustration_style\": \"frame\",\n            \"color_scheme\": \"black (#000000)\",\n            \"position\": \"surrounding the right-side image collage\",\n            \"special_instructions\": \"The border creates a distinct visual block for the imagery.\"\n          }\n        },\n        {\n          \"id\": \"image-4-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_976_105_2003_1038.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_976_105_2003_1038.jpg\"\n              }\n            ],\n            \"source\": \"Travel photography collage showing a woman at a monument and a couple walking with luggage\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, vertically centered\",\n            \"image_style\": \"No border on the photos themselves, but contained within the black frame mentioned in illustration-4-2\",\n            \"container\": \"The black frame (illustration-4-2)\",\n            \"special_instructions\": \"The image is a composite of two distinct photos placed side-by-side. The left photo shows a woman in a hat looking at a large archway. The right photo shows a couple walking away from the camera with a suitcase.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/d07410e8-bc20-4912-86d1-b92c47157eb3_input_0.jpg"], "target_image": "images/multi_image/d07410e8-bc20-4912-86d1-b92c47157eb3_target.jpg", "width": 2048, "height": 1152}
+{"index": 178, "id": "659f7644-5dfe-4b08-9ccf-9cd37d2f22fc", "task": "multi_image", "source": "real", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 5. CREATIVE DECORATIVE ELEMENTS (ENCOURAGED)\n*   **The Rule:** You are ENCOURAGED to add decorative visual elements beyond what is specified in the JSON blocks to enhance aesthetics.\n*   **Allowed Decorations:**\n    *   Decorative icons, symbols, or pictograms that relate to the slide content\n    *   Subtle geometric shapes, lines, or dividers\n    *   Background patterns, gradients, or texture overlays\n    *   Visual accents, highlights, or emphasis elements\n    *   Decorative borders, frames, or containers for content groups\n*   **Constraints:**\n    *   **DO NOT** add new text content - only use text from the JSON `content` fields\n    *   Decorative elements should complement the `global_style` theme\n    *   Keep decorations subtle and professional - they should enhance, not distract\n*   **Purpose:** Make each slide visually engaging and polished, going beyond a plain text-on-background layout.\n\n#### 6. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Wild Safari Minimalist\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#000000\",\n      \"secondary\": \"#444444\",\n      \"text\": \"#444444\",\n      \"accent\": null\n    },\n    \"special_instructions\": \"Clean, high-contrast design with a stark white background. The layout relies on the natural colors of the wildlife photography to provide visual interest rather than using colored graphical elements. Typography is bold and sans-serif for maximum readability.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Adventurous, Clean\",\n    \"composition_style\": \"Asymmetric, Two-Column\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 6,\n      \"blocks\": [\n        {\n          \"id\": \"title-6\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Days 2–3 – Tarangire National Park\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-right\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": null\n            },\n            \"container\": null,\n            \"special_instructions\": \"Large, bold font size. The text is left-aligned within the right-hand column of the slide.\"\n          }\n        },\n        {\n          \"id\": \"body-6-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Travel 3–4 hrs to Tarangire\\n• Game drives among elephants, lions, leopards, cheetahs\\n• Iconic baobab landscape\\n• Overnight at Tarangire camp\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#444444\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-right\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points using standard circular bullets. Text is left-aligned. Font weight is normal.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_113_114_671_1034.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_113_114_671_1034.jpg\"\n              }\n            ],\n            \"source\": \"Main feature image of a zebra\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left, vertically centered\",\n            \"image_style\": \"rounded corners\",\n            \"container\": null,\n            \"special_instructions\": \"Large vertical rectangular image on the left side. Has rounded corners. No border or shadow.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_720_710_1068_1036.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_720_710_1068_1036.jpg\"\n              }\n            ],\n            \"source\": \"Landscape image of clouds/mountains\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-center, left of center\",\n            \"image_style\": \"rounded corners\",\n            \"container\": null,\n            \"special_instructions\": \"Small square image. Has rounded corners. No border or shadow.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1120_711_1468_1036.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1120_711_1468_1036.jpg\"\n              }\n            ],\n            \"source\": \"Image of an antelope scratching itself\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-center, middle\",\n            \"image_style\": \"rounded corners\",\n            \"container\": null,\n            \"special_instructions\": \"Small square image. Has rounded corners. No border or shadow.\"\n          }\n        },\n        {\n          \"id\": \"figure-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1520_711_1866_1035.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1520_711_1866_1035.jpg\"\n              }\n            ],\n            \"source\": \"Image of giraffes by water\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-center, right of center\",\n            \"image_style\": \"rounded corners\",\n            \"container\": null,\n            \"special_instructions\": \"Small square image. Has rounded corners. No border or shadow.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\n\n**IMPORTANT: 4 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/659f7644-5dfe-4b08-9ccf-9cd37d2f22fc_input_0.jpg", "images/multi_image/659f7644-5dfe-4b08-9ccf-9cd37d2f22fc_input_1.jpg", "images/multi_image/659f7644-5dfe-4b08-9ccf-9cd37d2f22fc_input_2.jpg", "images/multi_image/659f7644-5dfe-4b08-9ccf-9cd37d2f22fc_input_3.jpg"], "target_image": "images/multi_image/659f7644-5dfe-4b08-9ccf-9cd37d2f22fc_target.jpg", "width": 2048, "height": 1152}
+{"index": 179, "id": "5f1d0187-f14c-4790-8b8c-d36fd7d3090d", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The slide features a clean, modern design with a light gray background and subtle, angled line art in the corners, echoing a technical diagram aesthetic. At the very top, centered, the title reads **\"HANDLING TOKEN UNCERTAINTY\"** in large, bold, dark navy sans‑serif type.\n\nBeneath the title, the slide is divided into two stacked sections, separated by a thin, soft-gray horizontal rule that spans most of the slide width.\n\n---\n\n### Top Section: AR Order\n\nOn the left side of the upper section is a small circular icon showing a simple forward arrow inside, symbolizing step-by-step progression. To the right of this icon, a compact heading appears in bold dark navy text:\n\n**\"AR ORDER\"**  \n*Confronts uncertainty directly*\n\nUnder this heading, a single concise line of body text in medium-gray reads:\n\n- **Sequential decoding meets difficult (forking) tokens as they appear.**\n\nTo the right of this short text block, occupying most of the upper-right quadrant, is the **sub image** placed prominently. The image is scaled so that its labels and colored nodes are clearly legible. It depicts the red “uncertain” token in the center of a chain of blue tokens, with arrows indicating forward generation in an autoregressive (AR) order, reinforcing the notion that uncertainty is encountered in sequence.\n\n---\n\n### Bottom Section: Arbitrary Order\n\nA second thin gray rule separates the lower section from the upper portion.\n\nAt the left of the bottom section, a circular icon containing a curved arrow represents reordering or skipping. Next to it, the heading appears in bold dark navy:\n\n**\"ARBITRARY ORDER\"**  \n*Bypasses uncertainty when needed*\n\nDirectly underneath, a brief line of text in medium-gray states:\n\n- **Model can skip hard tokens first, fill them once the context becomes easier.**\n\nBelow this text, aligned to the right and slightly overlapping the lower edge of the **sub image** (without obscuring it), a soft, semi-transparent annotation bar in light blue contains two short callouts in dark navy:\n\n- **\"Too hard? Bypass now.\"** positioned near the arc that skips the red token in the image.  \n- **\"Return when context is richer.\"** positioned near the arrow that loops back to the red token.\n\n---\n\nAt the very bottom center of the slide, in small, light-gray sans-serif text, a minimalist caption ties the visual together:\n\n**\"Visualizing how token order strategies confront or bypass uncertain positions during generation.\"**", "input_images": ["images/multi_image/5f1d0187-f14c-4790-8b8c-d36fd7d3090d_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 180, "id": "4013e573-5dfd-4b89-9757-40e7171ed9f2", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### OBJECTIVE  \nConvert the incoming Structured Content JSON into a single refined, presentation-ready 16:9 slide.\n\n### CONTENT RULES  \nDisplay only what is provided in the JSON. Do not invent labels, bullet points, taglines, or filler language. If the source text is brief, keep the layout intentionally minimal.\n\n### DESIGN GUIDELINES  \nTreat `global_style` as the master visual system. Follow its direction for background, color usage, typography, and overall mood. Apply any `special_instructions` precisely, especially those describing atmosphere, lighting, textures, or layout priorities.\n\n### BLOCK INTERPRETATION  \nRender blocks in line with their declared type, role, and hierarchy. Use `script` blocks purely as non-visible guidance for theme and emotion. Ensure text blocks have ample breathing room and remain easily legible.\n\n### IMAGE HANDLING  \nFor every block with `type: \"image\"`, you MUST place and display the referenced image files in the slide.\n\n### LAYOUT PRINCIPLES  \nCompose the slide with intentional structure. Allow variation: some slides may be dominated by imagery, others by text or simple structures. The finished layout should appear premium, cohesive, and suitable for a professional presentation.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Royal Corporate\",\n    \"background\": {\n      \"color\": \"#0E4D35\",\n      \"gradient\": \"radial-gradient(circle at center, #135C3F 0%, #0E4D35 100%)\",\n      \"texture\": \"none\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Times New Roman\",\n      \"secondary\": \"Times New Roman\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#D4AF37\",\n      \"secondary\": \"#F4E4BC\",\n      \"text\": \"#F4E4BC\",\n      \"accent\": \"#D4AF37\"\n    },\n    \"special_instructions\": \"The slide features a luxurious, formal aesthetic with a deep forest green background that has a subtle radial gradient lightening towards the center. All text and icons utilize a metallic gold color palette, ranging from a deep metallic gold (#D4AF37) for headings to a lighter cream-gold (#F4E4BC) for body text. The font is a classic serif typeface (Times New Roman) used in all caps for headings to convey authority and tradition. Icons are stylized with a metallic gold texture and subtle drop shadows to create depth.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Formal, Elegant, Authoritative\",\n    \"composition_style\": \"Centered Header, 2x2 Grid Layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 717,\n      \"blocks\": [\n        {\n          \"id\": \"illustration-717-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A stylized golden crown icon centered at the very top of the slide, acting as a decorative header element.\",\n          \"style\": {\n            \"illustration_style\": \"metallic icon\",\n            \"color_scheme\": \"gold\",\n            \"position\": \"top-center, above title\",\n            \"special_instructions\": \"Small size, centered horizontally.\"\n          }\n        },\n        {\n          \"id\": \"title-717\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"ENTERPRISE RISK FRAMEWORK\",\n          \"style\": {\n            \"color\": \"#D4AF37\",\n            \"font_family\": \"Times New Roman\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Text is in ALL CAPS with wide letter spacing.\"\n          }\n        },\n        {\n          \"id\": \"title-717-sub\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"CORE DOMAINS & GOVERNANCE FOCUS\",\n          \"style\": {\n            \"color\": \"#D4AF37\",\n            \"font_family\": \"Times New Roman\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center, below main title\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Text is in ALL CAPS, slightly smaller than main title.\"\n          }\n        },\n        {\n          \"id\": \"image-717-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_72_327_234_423.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_72_327_234_423.jpg\"\n              }\n            ],\n            \"source\": \"Definitions icon group\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"middle-left, top quadrant\",\n            \"image_style\": \"gold metallic texture\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Icon group showing an open book and a scroll document.\"\n          }\n        },\n        {\n          \"id\": \"body-717-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"RISK POLICY\\n• Unified risk appetite statement\\n• Clear ownership and escalation paths\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#F4E4BC\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left, top quadrant, right of icon\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Heading 'RISK POLICY' is in ALL CAPS and gold color. Bullet points are standard dots.\"\n          }\n        },\n        {\n          \"id\": \"illustration-717-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A golden icon group consisting of a stopwatch and a target with an arrow in the center.\",\n          \"style\": {\n            \"illustration_style\": \"metallic icon\",\n            \"color_scheme\": \"gold\",\n            \"position\": \"middle-right, top quadrant\",\n            \"special_instructions\": \"Represents time and targets.\"\n          }\n        },\n        {\n          \"id\": \"body-717-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"RISK IDENTIFICATION\\n• Structured scans across business units\\n• Prioritized register of critical exposures\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#F4E4BC\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-right, top quadrant, right of icon\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Heading 'RISK IDENTIFICATION' is in ALL CAPS and gold color.\"\n          }\n        },\n        {\n          \"id\": \"illustration-717-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A golden icon group consisting of a gear with an upward trend arrow and a magnifying glass.\",\n          \"style\": {\n            \"illustration_style\": \"metallic icon\",\n            \"color_scheme\": \"gold\",\n            \"position\": \"middle-left, bottom quadrant\",\n            \"special_instructions\": \"Represents monitoring and analysis.\"\n          }\n        },\n        {\n          \"id\": \"body-717-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"ASSESSMENT & CONTROLS\\n• Quantitative and qualitative scoring\\n• Mitigation plans with control owners\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#F4E4BC\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left, bottom quadrant, right of icon\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Heading 'ASSESSMENT & CONTROLS' is in ALL CAPS and gold color.\"\n          }\n        },\n        {\n          \"id\": \"illustration-717-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A golden icon group consisting of a document labeled 'CONTRACT' and a scale of justice.\",\n          \"style\": {\n            \"illustration_style\": \"metallic icon\",\n            \"color_scheme\": \"gold\",\n            \"position\": \"middle-right, bottom quadrant\",\n            \"special_instructions\": \"Represents legal terms and penalties.\"\n          }\n        },\n        {\n          \"id\": \"body-717-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"REPORTING & OVERSIGHT\\n• Board-level dashboards and trends\\n• Formal review cadence with key committees\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#F4E4BC\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-right, bottom quadrant, right of icon\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Heading 'REPORTING & OVERSIGHT' is in ALL CAPS and gold color.\"\n          }\n        },\n        {\n          \"id\": \"caption-717-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"CONFIDENTIAL – ENTERPRISE RISK MANAGEMENT OFFICE\",\n          \"style\": {\n            \"font_family\": \"Times New Roman\",\n            \"color\": \"#D4AF37\",\n            \"position\": \"bottom-center\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"special_instructions\": \"Small text, centered, with a thin horizontal gold line above it.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```", "input_images": ["images/multi_image/4013e573-5dfd-4b89-9757-40e7171ed9f2_input_0.jpg"], "target_image": "images/multi_image/4013e573-5dfd-4b89-9757-40e7171ed9f2_target.png", "width": 1376, "height": 768}
+{"index": 181, "id": "b5357b1e-df5d-4dd3-b15c-281c7907b1c4", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### MISSION\nTransform the input Structured Content JSON into one elegant, high-fidelity presentation slide.\n\n### STRICT CONTENT CONTROL\nOnly the JSON may supply visible content. Do not add explanatory copy, generated bullets, fake labels, placeholder text, or decorative words. If text is sparse, preserve the sparse feeling.\n\n### STYLE CONTROL\nHonor `global_style` as the governing design system. Use the specified background, color palette, typography, and overall visual mood. Apply `special_instructions` when provided, especially for atmosphere, lighting, texture, or layout emphasis.\n\n### BLOCK HANDLING\nRender visible blocks according to their role and hierarchy. Use `script` only as hidden context for the slide's theme and emotional tone. Keep text blocks readable and avoid cramped paragraphs.\n\n### IMAGE INTEGRITY & PRESERVATION\nWhen you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n\n### LAYOUT EXPECTATION\nDesign a slide with deliberate composition. Vary structure across slides: some may be visual-led, some text-led, some diagrammatic, some minimal. The output should feel premium, consistent, and ready for a professional deck.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Warm Afternoon Catnap\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#FDF7ED\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #FFE7C2 0%, #FDF7ED 55%, #F4E9DA 100%)\\\",\\n      \\\"texture\\\": \\\"paper\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Raleway\\\",\\n      \\\"secondary\\\": \\\"Lato\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#D9772C\\\",\\n      \\\"secondary\\\": \\\"#F4B97F\\\",\\n      \\\"text\\\": \\\"#5B4636\\\",\\n      \\\"accent\\\": \\\"#FFCF99\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"The slide should feel cozy and tranquil, echoing a warm sunlit afternoon. Use soft, rounded typography and generous white space. Subtle paper texture in the background, with warm beige and light cream tones that complement the orange tabby cat. Shadows are soft and diffuse, as if lit by natural window light. Corners of panels and image frames should be slightly rounded to enhance the gentle, relaxed mood.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Calm, Playful, Cozy\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 13,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-13\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"The Art of a Perfect Catnap\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#5B4636\\\",\\n            \\\"font_family\\\": \\\"Raleway\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, above the main image with generous margin\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"none\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Title in sentence case, slightly spaced lettering. Keep it large but leaving plenty of breathing room so the cat image remains the main focus.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/b5357b1e-df5d-4dd3-b15c-281c7907b1c4_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/b5357b1e-df5d-4dd3-b15c-281c7907b1c4_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"A relaxed orange tabby cat resting on a wooden floor with eyes gently closed.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning about 70% of slide width and anchored slightly below vertical center\\\",\\n            \\\"image_style\\\": \\\"soft drop shadow with subtle warm glow\\\",\\n            \\\"container\\\": \\\"rounded rectangle frame in soft cream\\\",\\n            \\\"special_instructions\\\": \\\"Apply a 16px border-radius to the image corners. Add a thin border in #F4B97F (2px). Shadow should be subtle, blurred, and offset downward to enhance the feeling of warmth and depth without overpowering the calm mood.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-13-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Slow down, stretch out, and let the quiet moments linger.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Lato\\\",\\n            \\\"color\\\": \\\"#7A5A41\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image frame\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Set in medium weight, small size compared to the title. Center-aligned line, with increased line spacing to keep it airy.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-13-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two soft, semi-transparent paw-print silhouettes placed diagonally opposite each other behind the corners of the image.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"minimal flat icon\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette with light #FFCF99 and faint #F4B97F\\\",\\n            \\\"position\\\": \\\"one behind upper-left of the image, another behind lower-right of the image\\\",\\n            \\\"special_instructions\\\": \\\"Keep opacity around 0.25 so the shapes are barely there, acting as gentle decorative accents that do not distract from the cat photo.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate a single widescreen slide from the JSON. The output should be refined, readable, and consistent with the specified visual system.", "input_images": ["images/multi_image/b5357b1e-df5d-4dd3-b15c-281c7907b1c4_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 182, "id": "a877df89-26ef-4e56-b7b1-5c496978d9b8", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide features a solid, deep charcoal gray background that serves as the canvas for a three‑column layout highlighting model performance comparison results. At the very top, centered across the full width, is the slide title **“Model Performance on Matching Benchmarks”** in a large, bold, white sans‑serif typeface. The body of the slide is divided into three equal vertical sections, separated by thin, vertical divider lines in a muted teal color that provide subtle structure without overwhelming the design.\n\nIn the **leftmost column**, the focus is **“Benchmark Overview”**. At the top of this section, placed prominently, is the sub image: a clear reproduction of the performance table showing methods in the leftmost column and three metric columns labeled *AUC-Hard (↑)*, *AUC-Medium (↑)*, and *AUC-Easy (↑)* with corresponding numeric scores. The table is displayed in its original orientation, occupying most of the upper half of this column. A faint, semi‑transparent starburst watermark is centered over the sub image, just light enough not to obstruct the numbers. Beneath the sub image, the heading **“Benchmark Overview”** appears in bold white text, followed by a compact explanatory line in smaller white font:  \n“Summarized AUC scores across difficulty levels, comparing classic and recent matching approaches.”\n\nThe **middle column** centers on **“Difficulty Levels”**. At the top of this column, instead of a separate photograph, the same sub image is re-used as a zoomed‑in, slightly enlarged crop emphasizing the three AUC columns labeled *Hard*, *Medium*, and *Easy*. This crop is aligned so that the column headers and a few representative score rows are clearly visible, while row labels at the far left are partially cropped to keep attention on the difficulty tiers. Again, a subtle, translucent starburst watermark overlays the center of this sub image. Below the image, the bold white heading **“Difficulty Levels”** is followed by a brief line in smaller white text:  \n“Compare model robustness as tasks move from hard to easy, focusing on AUC trends.”\n\nThe **rightmost column** is dedicated to **“Our Variants”**. The top of this section contains another usage of the same sub image, this time cropped to highlight only the bottom rows labeled “Ours (ELoFTR)” and “Ours (ROMA)” alongside their three AUC values. The crop is framed so these two methods and their scores are centered and easy to read, with the rest of the table fading out above them. As with the other columns, a faint starburst watermark sits atop the cropped sub image, keeping visual consistency across all three sections. Under this image, the heading **“Our Variants”** appears in bold white text, followed by a minimal explanatory sentence in smaller white type:  \n“Showcase performance of our ELoFTR‑ and ROMA‑based variants across benchmarks.”\n\nAcross the slide, margins are generous so that the sub image in each column has room to dominate visually while the text remains secondary and unobtrusive. All body text is left‑aligned within its column, yet the overall three‑column structure and the title are perfectly centered on the slide. The design is intentionally minimal, allowing the numerical details within the sub image to be the primary focus, while the restrained headings serve only to orient the viewer to what each view of the table emphasizes.", "input_images": ["images/multi_image/a877df89-26ef-4e56-b7b1-5c496978d9b8_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 183, "id": "ac8bdd30-9a7c-4532-a1a7-48207e70fad2", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### OBJECTIVE  \nConvert the input Structured Content JSON into a single polished, high-impact presentation slide.\n\n### CONTENT RULES  \nAll visible wording must come directly from the JSON. Do not invent extra labels, bullets, taglines, or filler phrases. If the JSON includes very little text, keep the design minimal to match.\n\n### VISUAL STYLE GUIDANCE  \nTreat `global_style` as the master design language. Follow its background, palette, typography, and overall mood. Apply any `special_instructions` precisely, including notes on atmosphere, lighting, textures, or layout accents.\n\n### BLOCK RENDERING LOGIC  \nDisplay each block according to its `role` and hierarchy in the JSON. Use `script` blocks, if any, only as invisible guidance for mood or narrative. Ensure all text blocks are comfortably legible, with ample spacing and no dense walls of text.\n\n### IMAGE HANDLING & PRESERVATION  \nFor every block where `type: \"image\"`, you MUST place the referenced source image on the slide. Do not crop away essential characters or labels.\n\n### LAYOUT EXPECTATIONS  \nCompose one widescreen slide with intentional structure. Respect the grid-based, three-column composition but allow variety in emphasis: some regions may be dominated by imagery, others by text, or remain airy and minimal. The final layout should appear refined, coherent, and suitable for a professional educational deck.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Educational Legal Infographic\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2E75B6\",\n      \"secondary\": \"#4CAF50\",\n      \"text\": \"#333333\",\n      \"accent\": \"#FF9800\"\n    },\n    \"special_instructions\": \"The slide uses a bright white canvas with a playful yet authoritative teaching vibe. A three-column grid in blue, green, and orange visually separates concepts. Cute cartoon animal mascots explain legal ideas in an accessible way, with subtle sparkles and soft shadows for depth but no clutter.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Educational, Playful, Professional\",\n    \"composition_style\": \"Grid-based, Three-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 601,\n      \"blocks\": [\n        {\n          \"id\": \"title-601\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"WHO IS THE SETTLOR IN A TRUST?\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, bold, uppercase heading centered along the top edge of the slide.\"\n          }\n        },\n        {\n          \"id\": \"subtitle-601\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"The Person Who Starts and Funds the Trust\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center, below main title\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Medium-size subtitle placed directly under the title.\"\n          }\n        },\n        {\n          \"id\": \"illustration-601-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A small pink rabbit mascot holding a sign that says 'SETTLOR' positioned to the left of the title.\",\n          \"style\": {\n            \"illustration_style\": \"cartoon character\",\n            \"color_scheme\": \"pink, beige, brown\",\n            \"position\": \"top-left, next to title\",\n            \"special_instructions\": \"Serves as a cute character introducing the topic.\"\n          }\n        },\n        {\n          \"id\": \"section-header-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"WHO THE SETTLOR IS\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, top\",\n            \"container\": \"blue rounded rectangle banner\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"White lettering on a solid blue rounded banner.\"\n          }\n        },\n        {\n          \"id\": \"body-601-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Sometimes called the Trustor or Grantor.\\n• The person or organization that creates the trust arrangement.\\n• Moves selected property out of personal ownership and into the trust.\\n• Sets the rules, goals, and conditions that govern the trust.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"minor\",\n            \"position\": \"left column, below header\",\n            \"container\": \"light blue rounded panel\",\n            \"container_opacity\": \"0.1\",\n            \"special_instructions\": \"Traditional bullet list using solid round bullets.\"\n          }\n        },\n        {\n          \"id\": \"image-601-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_52_14_150_182.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_52_14_150_182.jpg\"\n              }\n            ],\n            \"source\": \"Pink rabbit mascot holding a sign\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-left corner\",\n            \"image_style\": \"none\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Small mascot icon reinforcing the idea of the settlor.\"\n          }\n        },\n        {\n          \"id\": \"image-601-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_69_413_413_673.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_69_413_413_673.jpg\"\n              }\n            ],\n            \"source\": \"Blue rabbit transferring assets\",\n            \"caption\": \"MOVING ASSETS INTO THE TRUST\"\n          },\n          \"style\": {\n            \"position\": \"left column, bottom\",\n            \"image_style\": \"none\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Picture shows a blue rabbit placing coins and gems into a safe labeled 'THE TRUST'.\"\n          }\n        },\n        {\n          \"id\": \"section-header-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"WHAT THE SETTLOR DOES\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"center column, top\",\n            \"container\": \"green rounded rectangle banner\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"White text on a green rounded banner centered above the middle column.\"\n          }\n        },\n        {\n          \"id\": \"body-601-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"1. Sets Intent & Legal Terms: Uses a written trust deed to outline why the trust exists and how it must operate.\\n2. Provides Initial Assets: Places money, investments, or other property into the trust as starting capital.\\n3. Chooses Trustees: Appoints trusted people or institutions to manage the trust property.\\n4. Identifies Beneficiaries: Decides who can benefit from the trust (such as relatives or charities).\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"minor\",\n            \"position\": \"center column, below header\",\n            \"container\": \"light green rounded panel\",\n            \"container_opacity\": \"0.1\",\n            \"special_instructions\": \"Ordered list with numeric markers.\"\n          }\n        },\n        {\n          \"id\": \"image-601-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_485_455_892_707.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_485_455_892_707.jpg\"\n              }\n            ],\n            \"source\": \"Purple rabbit signing document\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center column, bottom\",\n            \"image_style\": \"none\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Shows a purple rabbit with glasses signing a trust document while other rabbits and children stand nearby.\"\n          }\n        },\n        {\n          \"id\": \"section-header-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"KEY POINTS TO WATCH\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, top\",\n            \"container\": \"orange rounded rectangle banner\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"White lettering on an orange rounded banner capping the right column.\"\n          }\n        },\n        {\n          \"id\": \"body-601-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"A. Tax Effects: Transfers into a trust can trigger tax consequences that depend on the trust type and local law.\\nB. Revocable or Irrevocable: With an irrevocable trust, the settlor normally cannot take assets back or change key terms.\\nC. Giving Up Control: To keep the trust effective, the settlor usually must hand day‑to‑day control of the assets to the trustees.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#333333\",\n            \"importance\": \"minor\",\n            \"position\": \"right column, below header\",\n            \"container\": \"light orange rounded panel\",\n            \"container_opacity\": \"0.1\",\n            \"special_instructions\": \"Alphabetical list using markers A, B, C.\"\n          }\n        },\n        {\n          \"id\": \"image-601-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_956_454_1296_696.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_956_454_1296_696.jpg\"\n              }\n            ],\n            \"source\": \"Orange rabbit with scales\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, bottom\",\n            \"image_style\": \"none\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Image of an orange rabbit weighing 'TAX & CONTROL' against 'BENEFITS & LEGALITY' on a balance scale.\"\n          }\n        },\n        {\n          \"id\": \"footer-601\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"TRUST LAW BASICS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#666666\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Small footer text with simple legal-themed icons (globe, document, building) arranged to its left.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```", "input_images": ["images/multi_image/ac8bdd30-9a7c-4532-a1a7-48207e70fad2_input_0.jpg", "images/multi_image/ac8bdd30-9a7c-4532-a1a7-48207e70fad2_input_1.jpg", "images/multi_image/ac8bdd30-9a7c-4532-a1a7-48207e70fad2_input_2.jpg", "images/multi_image/ac8bdd30-9a7c-4532-a1a7-48207e70fad2_input_3.jpg"], "target_image": "images/multi_image/ac8bdd30-9a7c-4532-a1a7-48207e70fad2_target.png", "width": 1376, "height": 768}
+{"index": 184, "id": "006cf029-b4c5-4e3b-b880-3c01feed5b3d", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, titled **“DETECTOR-FREE IMAGE MATCHING ARCHITECTURES”** in large gray 3D lettering at the top, is organized into three main horizontal sections that highlight different stages of a detector-free matching pipeline, with the sub image placed prominently on the right side as the central visual reference.\n\n- On the left, a narrow column labeled **“INPUT IMAGE PAIR”** briefly notes:  \n  *“Two related views are processed jointly, without explicit keypoint detectors.”*  \n  This text is minimal, leaving space for a small thumbnail of the sub image’s leftmost part to hint at the original pair of scenes.\n\n- The middle section, titled **“COARSE-TO-FINE FLOW ESTIMATION”**, runs across the slide. Immediately beside the text, the full sub image is inserted, occupying most of the slide area. Short captions around it point to the key blocks already visible in the sub image:  \n  - **“Coarse Encoders (CNN / DINO)”**  \n  - **“Transformer Flow Decoder”**  \n  - **“Multi-Scale Convolution Refiner”**  \n  - **“Self- & Cross-Attention Feature Transformation”**  \n  - **“Local Refinement to Fine Matches”**  \n  The annotations are kept concise so that the viewer’s focus remains on the architecture diagram in the sub image.\n\n- At the bottom, a slim footer titled **“DENSE & SEMI-DENSE MATCHING”** contains three short bullets aligned under the sub image:  \n  1. *Dense flow for pixel-wise correspondence.*  \n  2. *Semi-dense refinement for key regions.*  \n  3. *No handcrafted keypoint detectors required.*  \n\nThe overall layout centers the sub image as the primary element, with minimal supporting text to guide interpretation while maintaining a clean, visually driven slide.", "input_images": ["images/multi_image/006cf029-b4c5-4e3b-b880-3c01feed5b3d_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 185, "id": "87649599-cc48-4b6c-b17a-565f1dd16d62", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "Render a single presentation slide using the provided Structured Content JSON.\n\nTreat the JSON as the authoritative specification:\n1. Apply `global_style` to define the visual language of the slide.\n2. Follow `visual_hierarchy` to set emphasis, mood, and overall composition.\n3. Use `blocks` to determine the precise text and all visual elements.\n4. Honor `special_instructions` as the primary creative constraint whenever present.\n5. Use `source_images` only for embedding, scaling, and positioning within the layout.\n\nYou may not change, add, or remove any textual content from the JSON. All creativity must come from typographic and layout decisions only (e.g., line breaks, alignment, font size, spacing, weight, and placement).\n\nDesign the slide to look custom-crafted rather than templated. Avoid flat, repetitive bullet layouts. Use generous whitespace, logical grouping, tight alignment, and a composition that clearly communicates the step-by-step structure.\n\nMaintain the existing art direction: soft, welcoming, friendly-corporate. Use rounded shapes, the specified blue‑green gradient background, and bold, approachable sans-serif typography. Emphasize the numbered sequence of cards and ensure the flow arrows clearly guide the viewer.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Friendly Corporate Onboarding\",\n    \"background\": {\n      \"color\": \"#E0F7FA\",\n      \"gradient\": \"radial-gradient(circle at center, #E0F7FA, #B2EBF2)\",\n      \"texture\": \"none\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#4DB6AC\",\n      \"secondary\": \"#0277BD\",\n      \"text\": \"#263238\",\n      \"accent\": \"#FFB74D\"\n    },\n    \"special_instructions\": \"The slide features a soft, welcoming aesthetic with a light blue-green gradient background. The design uses rounded corners extensively on containers and buttons. The typography is bold and friendly, with a mix of sans-serif fonts. Visual hierarchy is established through numbered steps and distinct content cards.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Welcoming, Informative\",\n    \"composition_style\": \"Grid-based, Horizontal Flow\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 319,\n      \"blocks\": [\n        {\n          \"id\": \"title-319\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"ONBOARDING JOURNEY\",\n          \"style\": {\n            \"color\": \"#81D4FA\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"shadow\"\n              ],\n              \"outline_color\": \"#0277BD\",\n              \"shadow_type\": \"drop\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, bold title with a light blue fill and a darker blue outline, creating a 3D effect. Positioned at the top center of the slide.\"\n          }\n        },\n        {\n          \"id\": \"subtitle-319\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"WELCOME NEW TEAM MEMBERS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center, below main title\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, dark grey text, smaller than the main title but still prominent.\"\n          }\n        },\n        {\n          \"id\": \"card-1-319\",\n          \"type\": \"container\",\n          \"role\": \"body\",\n          \"content\": \"1 WELCOME & INTRODUCTION\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"supporting\",\n            \"position\": \"left, first in horizontal sequence\",\n            \"container\": \"rounded rectangle with light blue background (#E1F5FE)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Card contains a number '1' in a blue circle at the top left. Title is bold. Below the title is an image placeholder.\"\n          }\n        },\n        {\n          \"id\": \"image-1-319\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_0.jpg\",\n                \"relative_path\": \"images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_0.jpg\"\n              }\n            ],\n            \"source\": \"Welcome illustration\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"inside card 1, below title\",\n            \"image_style\": \"flat illustration style\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Illustration shows three diverse people waving and a handshake with a gift box.\"\n          }\n        },\n        {\n          \"id\": \"body-1-319\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"WELCOME & INTRODUCTION\\nGet to know the team & culture. Receive welcome kit.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"minor\",\n            \"position\": \"inside card 1, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Title repeated in bold, followed by two lines of descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"card-2-319\",\n          \"type\": \"container\",\n          \"role\": \"body\",\n          \"content\": \"2 COMPANY OVERVIEW\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"supporting\",\n            \"position\": \"second from left, in horizontal sequence\",\n            \"container\": \"rounded rectangle with light blue background (#E1F5FE)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Card contains a number '2' in a blue circle at the top left. Title is bold. Below the title is an image placeholder.\"\n          }\n        },\n        {\n          \"id\": \"image-2-319\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_1.jpg\",\n                \"relative_path\": \"images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_1.jpg\"\n              }\n            ],\n            \"source\": \"Company overview illustration\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"inside card 2, below title\",\n            \"image_style\": \"flat illustration style\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Illustration shows a building with a logo, a chart with an upward trend, and a circular logo.\"\n          }\n        },\n        {\n          \"id\": \"body-2-319\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"COMPANY OVERVIEW\\nMission, vision, & values. Business goals & structure.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"minor\",\n            \"position\": \"inside card 2, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Title repeated in bold, followed by two lines of descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"card-3-319\",\n          \"type\": \"container\",\n          \"role\": \"body\",\n          \"content\": \"3 ROLE & EXPECTATIONS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"supporting\",\n            \"position\": \"center, third in horizontal sequence\",\n            \"container\": \"rounded rectangle with light blue background (#E1F5FE)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Card contains a number '3' in a blue circle at the top left. Title is bold. Below the title is an image placeholder.\"\n          }\n        },\n        {\n          \"id\": \"image-3-319\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_2.jpg\",\n                \"relative_path\": \"images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_2.jpg\"\n              }\n            ],\n            \"source\": \"Role and expectations illustration\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"inside card 3, below title\",\n            \"image_style\": \"flat illustration style\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Illustration shows a person at a desk with a computer, surrounded by lightbulbs and documents.\"\n          }\n        },\n        {\n          \"id\": \"body-3-319\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"ROLE & EXPECTATIONS\\nJob responsibilities. Performance goals & metrics.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"minor\",\n            \"position\": \"inside card 3, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Title repeated in bold, followed by two lines of descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"card-4-319\",\n          \"type\": \"container\",\n          \"role\": \"body\",\n          \"content\": \"4 TOOLS & TECHNOLOGY\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"supporting\",\n            \"position\": \"fourth from left, in horizontal sequence\",\n            \"container\": \"rounded rectangle with light blue background (#E1F5FE)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Card contains a number '4' in a blue circle at the top left. Title is bold. Below the title is an image placeholder.\"\n          }\n        },\n        {\n          \"id\": \"image-4-319\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_3.jpg\",\n                \"relative_path\": \"images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_3.jpg\"\n              }\n            ],\n            \"source\": \"Tools and technology illustration\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"inside card 4, below title\",\n            \"image_style\": \"flat illustration style\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Illustration shows a laptop, tablet, phone, and various software icons like email, chat, and project management.\"\n          }\n        },\n        {\n          \"id\": \"body-4-319\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"TOOLS & TECHNOLOGY\\nSetting up accounts accounts & equipment. Learning internal software.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"minor\",\n            \"position\": \"inside card 4, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Title repeated in bold, followed by two lines of descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"card-5-319\",\n          \"type\": \"container\",\n          \"role\": \"body\",\n          \"content\": \"5 ONGOING SUPPORT\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"supporting\",\n            \"position\": \"top right, above card 6\",\n            \"container\": \"rounded rectangle with light blue background (#E1F5FE)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Card contains a number '5' in a blue circle at the top left. Title is bold. Below the title is an illustration placeholder.\"\n          }\n        },\n        {\n          \"id\": \"illustration-5-319\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Illustration of a person holding a certificate and a book, with trees in the background, representing training and development.\",\n          \"style\": {\n            \"illustration_style\": \"flat illustration\",\n            \"color_scheme\": \"green and blue tones\",\n            \"position\": \"inside card 5, below title\",\n            \"special_instructions\": \"Depicts learning and growth.\"\n          }\n        },\n        {\n          \"id\": \"body-5-319\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"TRAINING & DEVELOPMENT\\nComplete required courses. Explore learning opportunities.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"minor\",\n            \"position\": \"inside card 5, below illustration\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Title repeated in bold, followed by two lines of descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"card-6-319\",\n          \"type\": \"container\",\n          \"role\": \"body\",\n          \"content\": \"6 ONGOING SUPPORT\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom right, below card 5\",\n            \"container\": \"rounded rectangle with light blue background (#E1F5FE)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Card contains a number '6' in a blue circle at the top left. Title is bold. Below the title is an illustration placeholder.\"\n          }\n        },\n        {\n          \"id\": \"illustration-6-319\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Illustration of a manager talking to a group of employees with question marks above their heads, representing ongoing support and feedback.\",\n          \"style\": {\n            \"illustration_style\": \"flat illustration\",\n            \"color_scheme\": \"blue and yellow tones\",\n            \"position\": \"inside card 6, below title\",\n            \"special_instructions\": \"Depicts communication and problem-solving.\"\n          }\n        },\n        {\n          \"id\": \"body-6-319\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"ONGOING SUPPORT\\nCheck-ins with manager & team. Finding answers & feedback.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#263238\",\n            \"importance\": \"minor\",\n            \"position\": \"inside card 6, below illustration\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Title repeated in bold, followed by two lines of descriptive text.\"\n          }\n        },\n        {\n          \"id\": \"illustration-arrows-319\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A series of light blue, curved arrows connecting the cards horizontally from left to right, and then vertically from card 4 to card 5, and card 5 to card 6, indicating the flow of the onboarding journey.\",\n          \"style\": {\n            \"illustration_style\": \"flowchart connector\",\n            \"color_scheme\": \"light blue (#81D4FA)\",\n            \"position\": \"between and below the content cards\",\n            \"special_instructions\": \"Arrows guide the viewer's eye through the sequential steps.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\nRender this JSON into a single 16:9 slide that is visually engaging, easy to read, and faithful to all specified styling rules.", "input_images": ["images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_0.jpg", "images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_1.jpg", "images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_2.jpg", "images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_input_3.jpg"], "target_image": "images/multi_image/87649599-cc48-4b6c-b17a-565f1dd16d62_target.png", "width": 1376, "height": 768}
+{"index": 186, "id": "ea4cc598-5009-498a-9e64-1417fa21d512", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The presentation slide maintains a clean, professional layout on a white background, accented by a subtle, light grey geometric mesh pattern of interconnected dots and lines visible in the corners. At the top, a horizontal banner with a soft blue gradient and rounded corners frames the main title. The title reads \"SECURING THE PATH: FIVE KEY PHASES\" in a large, bold, dark blue sans-serif font, centered within the banner.\n\nThe main content is organized into a vertical flowchart located centrally on the slide. This diagram consists of five distinct nodes arranged in a column, connected by downward-pointing grey arrows to indicate a sequential process. The first node is a rounded square with a blue-to-green gradient containing a white cloud icon; to its right is the label \"Cloud Environment Setup\" in black text. The second node features a white shield icon within a similar gradient square, labeled \"Security Policies Defined\" to its right. The third node displays a white link chain icon, labeled \"Tunnel Establishment.\" The fourth node shows a white icon of a wireless router with two antennas, labeled \"Edge Gateway Integration.\" The final node at the bottom contains a white checkmark icon, labeled \"End-to-End Validation.\"\n\nTo the left of the central flowchart, three distinct blocks of text describe guidance for these stages, formatted in a black sans-serif font. Aligned with the top section, the text states: \"Begin with structured planning: confirm that subscriptions, credentials, and **allocated address spaces** are in place.\" The phrase \"allocated address spaces\" is bolded for emphasis. The second text block, positioned alongside the middle steps, reads: \"Define enforcement points and **path selection** to maintain protected **data flows**.\" Here, the words \"path selection\" and \"data flows\" are bolded. The final text block, aligned with the bottom steps, advises: \"**Confirm** and **observe** the secured path to maintain long-term stability,\" with the words \"Confirm\" and \"observe\" bolded.\n\nIn the bottom right corner of the slide, there is a decorative graphic element consisting of three flat icons arranged horizontally and separated by right-pointing arrows. From left to right, this sub-image displays a checklist clipboard, a yellow shield with a blue outline, and a blue wireless router emitting a signal.", "input_images": ["images/multi_image/ea4cc598-5009-498a-9e64-1417fa21d512_input_0.jpg", "images/multi_image/ea4cc598-5009-498a-9e64-1417fa21d512_input_1.jpg"], "target_image": "images/multi_image/ea4cc598-5009-498a-9e64-1417fa21d512_target.png", "width": 1376, "height": 768}
+{"index": 187, "id": "738b1b21-6de3-4908-b245-31e77f1b923a", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Designing an ethereal fantasy art presentation slide themed around a “Bohemian Forest Muse,” this composition highlights watercolor textures, ornate jewelry, and floral headpieces inspired by the sub image. The slide background is clean, plain white to emphasize the artwork.\n\nAt the top left, a wide, deep teal rectangular banner with a thin gold border contains the title text “BOHEMIAN FOREST MUSE” in large, hand‑painted style capital letters. The letters are dark teal with soft gold outlines, and delicate ink splatters radiate gently from the banner’s edges.\n\nBeneath this banner, aligned along the left side of the slide, three rectangular sub‑images are stacked vertically. Each one is framed with a very thin, soft‑gold border:\n1. The first sub‑image crops the upper portion of the sub image, focusing on the feathered headdress and blooming flowers around the hair.\n2. The second sub‑image centers on the jewelry: the layered necklaces, the large circular earrings, and the detailed pendant work.\n3. The third sub‑image highlights the flowing clothing and color gradients, emphasizing the transition between warm oranges and cool blues.\n\nTo the right of these stacked sub‑images, the slide is divided into a slim central column of minimal text and a large right‑hand canvas reserved for the main art.\n\nThe central text column is broken into three concise sections, all using a simple black sans‑serif typeface:\n\n• The first section is titled “COLOR AURA” in bold uppercase. Under it, two short bullet points, each with a small teal circular bullet:\n  – “Warm florals & cool shadows”\n  – “Watercolor splashes”\n\n• The second section is titled “ORNAMENT & DETAIL,” followed by:\n  – “Layered jewelry”\n  – “Textured fabrics”\n\n• The third section reads “MOOD & STORY,” listing:\n  – “Mystic forest spirit”\n  – “Quiet, dreamlike gaze”\n\nOn the right half of the slide, dominating the layout, the sub image is placed large and prominent. It is centered vertically and horizontally within its area, surrounded by a subtle, soft‑edge drop shadow to lift it from the white background. No additional frames or elements cross over this region; the artwork remains the main focus. The vivid feathers, roses, and cascading hair from the sub image occupy most of this section, with the blurred face area preserved exactly as in the original.\n\nIn the lower right corner, beneath the sub image, a small, light‑grey caption in unobtrusive text reads: “Sub image: Bohemian figure in watercolor bloom.”\n\nThe overall slide design keeps text minimal and secondary, allowing the sub image to command attention while the clean white background and restrained teal‑and‑gold accents echo the image’s rich yet airy fantasy atmosphere.", "input_images": ["images/multi_image/738b1b21-6de3-4908-b245-31e77f1b923a_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 188, "id": "2e1d9952-6826-4be4-a0ae-6f1f72b1df97", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Organized into a four-quadrant grid layout on a white background, this presentation slide titled **\"LANDSCAPE OF THE GREAT WALL\"** uses soft, pastel tones and hand-drawn style accents to highlight a serene mountain scene. At the top center, the main title **\"LANDSCAPE OF THE GREAT WALL\"** is displayed in large, bold, black, uppercase letters. To the right of the title, two small gold stars and a tiny sketch of a mountain outline add a subtle decorative touch. The slide is divided into four rounded-rectangle sections, each defined by a distinct colored border, while the sub image is the visual anchor of the entire composition.\n\n---\n\n### Top-Left Quadrant – Blue Border  \n- Header in black uppercase letters inside a light-blue label: **\"DAWN OVER RIDGES\"**.  \n- Below the header, a short caption in small black text reads: **\"Soft light, layered hills, quiet air.\"**  \n- The sub image is placed here, occupying most of the quadrant’s width.  \n  - The wall runs diagonally from the lower left toward the center, its stone parapets catching the warm sun.  \n  - Misty mountain ranges recede into the background, fading from dark foreground greens to pale blues and creams.  \n  - The sun glows near the upper-right edge of the sub image, casting a hazy, golden atmosphere.  \n- Around the image, hand-drawn blue contour lines echo the shapes of the hills, reinforcing the sense of depth.\n\n---\n\n### Top-Right Quadrant – Red Border  \n- Header in a pale-red label: **\"LINES OF STONE\"**.  \n- Underneath, minimal descriptive text in small type:  \n  - **\"Ancient walls tracing the spine of the mountains.\"**  \n- A reduced, sketch-like inset of the sub image appears on the right side of this quadrant, simplified to outlines:  \n  - The wall is represented as a single continuous zigzag line along stylized peaks.  \n  - Tiny rectangular towers punctuate the line, suggesting watchtowers.  \n- To the left of this sketch, three bullet-like symbols (small red circles) each with a single word beside them:  \n  - **\"Height\"**  \n  - **\"Distance\"**  \n  - **\"Silence\"**  \n- Faint, red, upward-slanting arrows behind the sketch hint at the climb along the wall without overwhelming the visuals.\n\n---\n\n### Bottom-Left Quadrant – Orange Border  \n- Header in a soft-orange label: **\"MIST & LIGHT\"**.  \n- Centered beneath the header is a cropped, panoramic strip taken from the upper half of the sub image:  \n  - Only the overlapping silhouettes of the mountain ridges and the glowing sky are shown.  \n  - The gradation of tones—from deeper gold near the sun to pale, smoky gray farther away—is preserved.  \n- Below this strip, a single line of text in thin black letters:  \n  - **\"Layers of haze soften every edge.\"**  \n- Light orange arcs and dots extend outward from the sun’s position in the strip, resembling hand-drawn rays, subtly emphasizing the illumination.\n\n---\n\n### Bottom-Right Quadrant – Green Border  \n- Header in a pale-green label: **\"PATHS THROUGH TIME\"**.  \n- The main feature is another placement of the sub image, slightly reduced and aligned to the left of the quadrant:  \n  - Here, the lower portion is emphasized, showing the autumn-colored vegetation hugging the wall’s base.  \n  - The wall’s curve guides the eye from foreground to background, suggesting movement and journey.  \n- To the right of this image, three short, stacked phrases in bold black text, each on its own line:  \n  - **\"Stone steps.\"**  \n  - **\"Mountain winds.\"**  \n  - **\"Endless horizon.\"**  \n- A faint, hand-drawn green path line sweeps from the lower-left corner of the quadrant up toward the text, visually connecting the viewer’s eye from the base of the wall toward the sky.\n\n---\n\nAcross the slide, all text remains sparse and secondary, ensuring the sub image of the Great Wall and its surrounding mountains is the primary focus. White space around each quadrant keeps the composition airy and uncluttered, allowing the viewer to linger on the atmospheric landscape while the minimal captions quietly guide attention to dawn, structure, mist, and the sense of timeless passage.", "input_images": ["images/multi_image/2e1d9952-6826-4be4-a0ae-6f1f72b1df97_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 189, "id": "546d9e05-81fc-4bea-9648-5096c8ca58d1", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "The first image features a close-up of a mint-green and white striped tiger character holding a candy cane, which is then placed on the left side of the slide. The slide itself, titled \"WELCOME MISTY MINT!\" in large red and green lettering at the top, introduces the character with the subtitle \"MISTY MINT: The Frosty Trendsetter.\" The content is organized into four sections: \"PERSONALITY: Gearlyn: Cheerful, Bubbly, Polite, Daydreamy,\" followed by \"LOOKS & TRAITS: Species: Tiger Gender: Female Fur: Mint Green & White Stripes Eyes: Big Brown Esea: Color Style: Cozy/Cute: Mint Green & Candy Cane Red Hobby: Decorating/Baking.\" The next section, \"HOME & LIFESTYLE,\" reads \"House Style: Nordic cottage with a winter-mint theme. Furniture: Pastel wood pieces, peppermint cushions, fairy lights, potted herbs. Favorite Song: Café K.K. Phrase: 'Minty-fresh!'\". The final section states \"FUN FACT: Famous for her cozy cocoa nights and cookie-baking marathons! Loves gifting handmade scarves and sweets.\" Decorative icons like leaves, stars, and a house accent the text on the right.", "input_images": ["images/multi_image/546d9e05-81fc-4bea-9648-5096c8ca58d1_input_0.jpg"], "target_image": "images/multi_image/546d9e05-81fc-4bea-9648-5096c8ca58d1_target.png", "width": 1376, "height": 768}
+{"index": 190, "id": "ea864520-722b-4063-b5d2-b5d8d8b8ff40", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Showcasing an educational infographic titled \"REMOTE MONITORING CYCLE\" designed with a playful, hand-drawn cartoon aesthetic against a light off-white background. The title is rendered in large, 3D block letters with a teal face and coral sides, flanked by small gear icons in yellow and purple on the left, and turquoise and red on the right, interspersed with four-pointed sparkles. The slide is structured horizontally into four distinct stages, each contained within a rounded rectangular frame with a thick black border. Between these frames, dotted arrows indicate the progression from left to right.\n\nThe first stage on the far left is headed by the text \"1. CAPTURE & SENSING\" in bold, black, uppercase letters. Below this header sits the first sub-image, depicting a blue industrial machine with pipes and three gauges; one gauge features a thermometer, while the other two have smiling faces. Underneath this visual, the section title \"1. CAPTURE & SENSING\" is repeated in bold, black font. Below that is a paragraph of body text in a smaller, standard sans-serif font: \"INSTRUMENTS CONTINUOUSLY TRACK CONDITIONS SUCH AS HEAT, PRESSURE, MOTION, AND FLOW. RAW READINGS ARE TURNED INTO DIGITAL OR ANALOG ELECTRICAL SIGNALS FOR FURTHER USE.\"\n\nMoving to the second section, the header reads \"2. LINK & TRANSFER\" in the same bold style. The corresponding sub-image shows a communication tower with a smiling orange face emitting radio waves, connected by a thick cable to a satellite dish. Lightning bolts labeled \"WIRELESS\" and a cable labeled \"WIRED\" illustrate the transmission methods. Beneath this image, the title \"2. LINK & TRANSFER\" is repeated, followed by the description: \"MEASURED SIGNALS TRAVEL THROUGH NETWORK PATHS TO A CENTRAL LOCATION. TECHNOLOGIES CAN INCLUDE RF LINKS, SATELLITE OR CELLULAR SYSTEMS, AND DEDICATED PHYSICAL CABLES.\"\n\nThe third section features the header \"3. ANALYSIS & VISUALS\". The sub-image displays a control room scene where a person wearing a headset operates a keyboard in front of a wall of monitors showing graphs, a world map, and numerical data. The text below the image reads \"3. ANALYSIS & VISUALS\" in bold, followed by the explanatory text: \"INCOMING STREAMS ARE FILTERED, STORED, AND COMPUTED INTO MEANINGFUL METRICS. INTERACTIVE SCREENS, MAPS, AND CHARTS HIGHLIGHT TRENDS AND ALERT OPERATORS TO CHANGES.\"\n\nThe final section on the right is titled \"4. RESPONSE & CONTROL\". The sub-image illustrates a rocket launch scene with a worker adjusting a control panel in the foreground and another worker tending to machinery in the background, with green arrows indicating feedback loops. The bold text \"4. RESPONSE & CONTROL\" appears below the image, followed by the final descriptive paragraph: \"INSIGHTS DRIVE AUTOMATED CONTROLS AND HUMAN INTERVENTIONS, SUPPORTING RAPID CORRECTIONS AND PERFORMANCE TUNING. TYPICAL USE CASES INCLUDE INDUSTRIAL OPERATIONS, TRANSPORTATION SYSTEMS, AND EXPERIMENTAL TEST FACILITIES.\" The entire layout uses a consistent color palette of muted blues, oranges, greens, and greys, maintaining a cohesive and engaging visual style throughout.", "input_images": ["images/multi_image/ea864520-722b-4063-b5d2-b5d8d8b8ff40_input_0.jpg", "images/multi_image/ea864520-722b-4063-b5d2-b5d8d8b8ff40_input_1.jpg", "images/multi_image/ea864520-722b-4063-b5d2-b5d8d8b8ff40_input_2.jpg", "images/multi_image/ea864520-722b-4063-b5d2-b5d8d8b8ff40_input_3.jpg"], "target_image": "images/multi_image/ea864520-722b-4063-b5d2-b5d8d8b8ff40_target.png", "width": 1376, "height": 768}
+{"index": 191, "id": "ecc80136-a7fe-4e9d-814d-c074f5409595", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM\nYou are a narrative slide compositor with rigorous visual communication standards.\n\n### OBJECTIVE\nRender a single, production-ready slide image from the input JSON, preserving semantic accuracy while maintaining deck-level stylistic coherence.\n\n### RENDERING POLICY\n- Only display textual elements that originate from JSON `content` fields; do not invent or paraphrase any copy.\n- `global_style` governs the complete visual system: background treatment, color scheme, typography, texture, dimensionality, and spatial rhythm.\n- `special_instructions` may refine the art direction, including atmosphere, materials, camera angle, or ornamental details.\n- Enforce a clear typographic hierarchy: titles must dominate, body text must remain comfortably readable, captions should be understated, and visuals must never obscure key text.\n- Use `source_images` exclusively for imagery placed on the slide.\n\n### DESIGN DIRECTION\nSelect a layout that feels distinctive yet consistent with the defined style—options include a bold hero setup, high-end editorial spread, elegant data visual, modular card arrangement, central statement layout, or an off-center grid. Ensure the final slide is sharp, legible, visually balanced, and free from any fabricated elements.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Cultural Heritage Earth Tones\",\n    \"background\": {\n      \"color\": \"#D6B888\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Georgia\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2F3E35\",\n      \"secondary\": \"#F2E9D8\",\n      \"text\": \"#3E2723\",\n      \"accent\": \"#8D6E63\"\n    },\n    \"special_instructions\": \"The slide features a warm, earthy aesthetic with a solid beige/tan background. The design uses a serif font for the title to evoke a traditional or academic feel, contrasted with a clean sans-serif for body text. Decorative elements include subtle line art icons in the corners (compasses, books) and a large, colorful circular illustration on the right. The overall mood is educational, respectful, and culturally rich.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Educational, Cultural, Professional\",\n    \"composition_style\": \"Two-column (Text Left, Image Right)\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 11,\n      \"blocks\": [\n        {\n          \"id\": \"title-11\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Living Traditions: Language, Arts, and Shared Learning\",\n          \"style\": {\n            \"color\": \"#2F3E35\",\n            \"font_family\": \"Georgia\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"shadow\"\n              ],\n              \"outline_color\": null,\n              \"shadow_type\": \"drop\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large serif title, centered at the top with a soft drop shadow to lift it from the background.\"\n          }\n        },\n        {\n          \"id\": \"body-11-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Local languages, music, and stories remain central to community identity.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3E2723\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, upper half\",\n            \"container\": \"rounded rectangle panel with light beige background (#F2E9D8)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Left-aligned text inside a rounded rectangle with a fine border, leaving generous padding on all sides.\"\n          }\n        },\n        {\n          \"id\": \"body-11-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Schools and cultural spaces collaborate to celebrate diverse voices and traditions.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3E2723\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, lower half\",\n            \"container\": \"rounded rectangle panel with light beige background (#F2E9D8)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Left-aligned text within a rounded rectangle matching the panel above, aligned to create a clean vertical stack.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_817_238_1265_684.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_817_238_1265_684.jpg\"\n              }\n            ],\n            \"source\": \"Cultural diversity illustration\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, vertically centered\",\n            \"image_style\": \"circular composition with flat vector art style\",\n            \"container\": \"none\",\n            \"special_instructions\": \"A large circular collage containing diverse faces, musical instruments (sitar, violin, guitar), open books, and various script characters (Devanagari, Gurmukhi, Chinese, Arabic) arranged in a dense, colorful pattern.\"\n          }\n        },\n        {\n          \"id\": \"illustration-11-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Fine line-art symbols in the slide corners: an open book in the top-left, a compass in the top-right, and complementary academic icons in the bottom corners.\",\n          \"style\": {\n            \"illustration_style\": \"line art\",\n            \"color_scheme\": \"muted brown/grey matching the theme\",\n            \"position\": \"four corners of the slide\",\n            \"special_instructions\": \"Use these as understated framing elements that support the educational and exploratory theme without distracting from the main content.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nCreate the final slide in 16:9 aspect ratio. Keep the look premium and varied, and ensure every visible element is justified by the JSON.", "input_images": ["images/multi_image/ecc80136-a7fe-4e9d-814d-c074f5409595_input_0.jpg"], "target_image": "images/multi_image/ecc80136-a7fe-4e9d-814d-c074f5409595_target.png", "width": 1376, "height": 768}
+{"index": 192, "id": "799cfc44-4ff5-4dd9-954a-5ed9e6c3b8d7", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### OBJECTIVE  \nTurn the provided Structured Content JSON into a polished, high-impact presentation slide.\n\n### CONTENT RULES  \nAll visible words must come directly from the JSON. Do not invent bullets, headings, labels, annotations, or filler phrases. If the JSON is minimal, keep the layout minimal.\n\n### VISUAL STYLE GUIDANCE  \nTreat `global_style` as the master design language. Follow its background, palette, typography, and mood directions. When `special_instructions` appear, prioritize them for atmosphere, lighting, textures, and layout cues.\n\n### BLOCK RENDERING LOGIC  \nDisplay each block according to its `role` and structural importance. Use any `script` blocks solely as backstage guidance for theme and feeling, never as visible copy. Ensure text blocks remain legible with comfortable spacing.\n\n### IMAGE HANDLING  \nFor any block with `type: \"image\"`, you MUST place the referenced image files in the slide exactly as given. Do not replace, crop away key elements, or swap in alternative art.\n\n### COMPOSITION EXPECTATIONS  \nConstruct a single widescreen slide with intentional layout. Let composition vary naturally—some slides may emphasize imagery, others text, diagrams, or empty space. The final result should appear cohesive, premium, and suitable for professional use.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Memphis Pop Art\",\n    \"background\": {\n      \"color\": \"#3B82F6\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": \"Scattered geometric shapes (triangles, zig-zags, squiggles) and halftone dots in pink, yellow, and black\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#FF6B9D\",\n      \"secondary\": \"#FCD34D\",\n      \"text\": \"#000000\",\n      \"accent\": \"#10B981\"\n    },\n    \"special_instructions\": \"The slide features a vibrant, high-energy Memphis design style. It uses bold, thick black outlines for all text and shapes. The color palette is a mix of bright pastels and saturated primary colors (pink, yellow, teal, orange) against a bright blue background. Decorative elements include geometric patterns like chevrons, grids, and wavy lines.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Playful, Energetic, Creative\",\n    \"composition_style\": \"Grid-based with a dominant central illustration and surrounding text blocks\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 305,\n      \"blocks\": [\n        {\n          \"id\": \"title-305\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"CREATIVE AGILITY\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"outline\",\n                \"shadow\"\n              ],\n              \"outline_color\": \"#000000\",\n              \"shadow_type\": \"drop\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, bold, uppercase text with a thick black outline and a slight drop shadow. The letters have alternating fill colors: 'CREATIVE' is pink, 'AGILITY' is yellow.\"\n          }\n        },\n        {\n          \"id\": \"image-305-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_87_138_891_449.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_87_138_891_449.jpg\"\n              }\n            ],\n            \"source\": \"Central illustration depicting a character holding an 'ADAPT' sign amidst chaotic arrows, gears, and question marks\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center, spanning horizontally across the middle of the slide\",\n            \"image_style\": \"flat vector illustration with thick black outlines\",\n            \"container\": \"none\",\n            \"special_instructions\": \"The illustration is the visual anchor of the slide, placed behind the title but above the bottom text blocks.\"\n          }\n        },\n        {\n          \"id\": \"body-305-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"1. EXPERIMENTAL MINDSET\\n• Treat every challenge as a prototype.\\n• Iterate quickly instead of waiting for perfect answers.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left\",\n            \"container\": \"rectangular box with a pink chevron pattern header and white body\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a white box with a thick black border. The header background is pink with white zig-zag lines. The text is left-aligned with bullet points.\"\n          }\n        },\n        {\n          \"id\": \"body-305-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"2. DYNAMIC PROBLEM SOLVING\\n• Reframe obstacles from multiple angles.\\n• Combine unusual ideas to unlock fresh options.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-center\",\n            \"container\": \"rectangular box with a yellow grid pattern header and teal body\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a teal box with a thick black border. The header background is yellow with a white grid pattern. The text is left-aligned with bullet points.\"\n          }\n        },\n        {\n          \"id\": \"body-305-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"3. RESILIENT MOMENTUM\\n• Move forward even when conditions shift.\\n• Learn fast from missteps and reset direction.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right\",\n            \"container\": \"rectangular box with an orange wavy pattern header and white body\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a white box with a thick black border. The header background is orange with blue wavy lines. The text is left-aligned with bullet points.\"\n          }\n        },\n        {\n          \"id\": \"body-305-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"4. NAVIGATING OPTIONS\\n• Scan multiple paths before committing.\\n• Choose a direction, then refine as you go.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-right\",\n            \"container\": \"rectangular box with a blue header and pink chevron pattern side\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a white box with a thick black border. The header background is blue. The right side of the header has a pink chevron pattern. The text is left-aligned with bullet points.\"\n          }\n        },\n        {\n          \"id\": \"caption-305-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"MEMPHIS STYLE WORKSHOP\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"bottom-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"special_instructions\": \"Small, uppercase text centered at the very bottom of the slide.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nRender one widescreen slide from this JSON, ensuring clarity, balance, and adherence to the defined visual language.", "input_images": ["images/multi_image/799cfc44-4ff5-4dd9-954a-5ed9e6c3b8d7_input_0.jpg"], "target_image": "images/multi_image/799cfc44-4ff5-4dd9-954a-5ed9e6c3b8d7_target.png", "width": 1376, "height": 768}
+{"index": 193, "id": "aa79bda3-6f46-4573-a9af-66be9f2dd212", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM\nYou are a cinematic presentation renderer with strong graphic design discipline.\n\n### OBJECTIVE\nGenerate a single polished slide image from the input JSON while preserving both semantic accuracy and deck-level style consistency.\n\n### RENDERING POLICY\n- Text fidelity is absolute: only render text from JSON `content` fields.\n- `global_style` controls the full visual language, including background, palette, fonts, texture, depth, lighting, and spacing.\n- `special_instructions` can introduce specific art direction such as atmosphere, materials, camera feel, or decorative treatment.\n- Maintain clear hierarchy: titles should lead, supporting text should be readable, captions should remain subtle, and visuals should not obscure essential text.\n- Use `source_images` to embed in the slide.\n\n### DESIGN DIRECTION\nChoose a layout that makes the slide memorable while staying coherent with the given style: bold hero composition, premium editorial layout, refined infographic, modular card layout, centered statement, or asymmetric grid. Keep the final output crisp, legible, balanced, and free of invented content.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Modern Grace Reading Moment\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#F7F3F0\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #FFFFFF 0%, #F7F3F0 45%, #E4DAD3 100%)\\\",\\n      \\\"texture\\\": \\\"paper\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Playfair Display\\\",\\n      \\\"secondary\\\": \\\"Source Sans Pro\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#E24991\\\",\\n      \\\"secondary\\\": \\\"#B3C0CD\\\",\\n      \\\"text\\\": \\\"#333333\\\",\\n      \\\"accent\\\": \\\"#F6A623\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall feel is elegant, calm, and editorial, inspired by a lifestyle magazine spread. Soft warm lighting, subtle paper texture, and generous white space surround the main image. Accents pick up the fuchsia dress and muted grey blazer from the photo. No harsh lines; use gentle rounded corners and soft drop shadows to keep the layout sophisticated and light.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Elegant, Professional, Calm\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 29,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-29\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Ageless Style, Quiet Confidence\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#333333\\\",\\n            \\\"font_family\\\": \\\"Playfair Display\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, slightly above the main image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"none\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"none\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Set in large serif type with generous letter spacing, on a single line or two balanced lines, centered.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/aa79bda3-6f46-4573-a9af-66be9f2dd212_input_0.png\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/aa79bda3-6f46-4573-a9af-66be9f2dd212_input_0.png\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Elegant older woman in a grey blazer and pink dress reading a magazine at a cafe table.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center of the slide, spanning about 70% of the slide width and vertically dominant\\\",\\n            \\\"image_style\\\": \\\"subtle soft-shadow with rounded corners\\\",\\n            \\\"container\\\": \\\"Thin off-white frame with slight inner padding\\\",\\n            \\\"special_instructions\\\": \\\"Apply a 16–20px border-radius to the image, with a soft drop shadow (#000000 at 12% opacity, blurred). Keep the frame minimal so the photo remains the clear focal point.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-29-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Curated living, one thoughtful page at a time.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Source Sans Pro\\\",\\n            \\\"color\\\": \\\"#555555\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image frame\\\",\\n            \\\"container\\\": null,\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Small, centered body text with relaxed line spacing; keep it understated so it supports rather than competes with the image.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-29-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two overlapping soft gradient circles, one blush pink and one muted grey, subtly peeking from behind the lower-right corner of the image.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette (#E24991 and #B3C0CD)\\\",\\n            \\\"position\\\": \\\"behind and slightly offset at the lower-right of the main image\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (around 25–30%), large radius circles that create a gentle layered effect without distracting from the photo; partially cropped by the slide edge for a modern editorial feel.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate the slide image as a clean widescreen presentation frame. Respect `global_style`, `visual_hierarchy`, `blocks`, and any `special_instructions`.", "input_images": ["images/multi_image/aa79bda3-6f46-4573-a9af-66be9f2dd212_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 194, "id": "4a5fdbe4-f899-4648-bf0e-797174efd9bb", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are an AI slide-rendering system. Transform the provided slide JSON into a polished presentation graphic.\n\nApply these rules precisely:\n- Treat the JSON as the single source of truth. Display only the text contained in `blocks`; do not invent or omit wording.\n- Place every referenced image from the JSON onto the slide.\n- Use `global_style` as the core visual identity. Keep its typography, palette, background, and overall aesthetic intact.\n- Use `visual_hierarchy` to guide prominence so that key information appears visually dominant and supporting details remain subtle.\n- When present, `special_instructions` may adjust or supersede default styling; follow them rigorously.\n\nCompose the layout with sophistication. Avoid relying on a generic title-plus-bullets arrangement. Employ asymmetrical layouts, column structures, modular panels, prominent visuals, short annotations, and deliberate empty space where suitable. The finished slide must appear refined, legible, and ready for formal presentation.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Elegant Technical Presentation\",\n    \"background\": {\n      \"color\": \"#F5E6D3\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Times New Roman\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2C3E50\",\n      \"secondary\": \"#8B4513\",\n      \"text\": \"#1A1A1A\",\n      \"accent\": \"#5D4037\"\n    },\n    \"special_instructions\": \"The slide features a sophisticated, academic aesthetic with a warm beige background. It uses a serif font for the main title to convey authority and a sans-serif font for body text for readability. Decorative elements include cherry blossom branches in the corners and bamboo on the left, adding an artistic, natural touch to the technical content. The layout is structured with a clear two-column division for the main content area.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Technical, Elegant\",\n    \"composition_style\": \"Two-column with decorative borders\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 741,\n      \"blocks\": [\n        {\n          \"id\": \"title-741\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"DEEP EXCAVATION SUPPORT: TANGENT PILE SYSTEM\",\n          \"style\": {\n            \"color\": \"#1A1A1A\",\n            \"font_family\": \"Times New Roman\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, centered title in all caps with a classic serif font.\"\n          }\n        },\n        {\n          \"id\": \"illustration-741-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative cherry blossom branch with pink flowers extending from the top-right corner downwards.\",\n          \"style\": {\n            \"illustration_style\": \"watercolor illustration\",\n            \"color_scheme\": \"pink, brown, green\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"Artistic, natural element framing the slide content.\"\n          }\n        },\n        {\n          \"id\": \"illustration-741-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative cherry blossom branch with pink flowers extending from the top-left corner downwards.\",\n          \"style\": {\n            \"illustration_style\": \"watercolor illustration\",\n            \"color_scheme\": \"pink, brown, green\",\n            \"position\": \"top-left corner\",\n            \"special_instructions\": \"Artistic, natural element framing the slide content.\"\n          }\n        },\n        {\n          \"id\": \"illustration-741-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative bamboo stalks with green leaves extending from the bottom-left corner upwards.\",\n          \"style\": {\n            \"illustration_style\": \"watercolor illustration\",\n            \"color_scheme\": \"green, brown\",\n            \"position\": \"bottom-left corner\",\n            \"special_instructions\": \"Artistic, natural element framing the slide content.\"\n          }\n        },\n        {\n          \"id\": \"illustration-741-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative bamboo leaves extending from the bottom-right corner upwards.\",\n          \"style\": {\n            \"illustration_style\": \"watercolor illustration\",\n            \"color_scheme\": \"green, brown\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Artistic, natural element framing the slide content.\"\n          }\n        },\n        {\n          \"id\": \"body-741-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"GEOMETRY & SOIL PROFILE\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, top\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Section header in bold, all caps, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"figure-741-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_41_217_615_721.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_41_217_615_721.jpg\"\n              }\n            ],\n            \"source\": \"Cross-section diagram of tangent pile walls showing soil layers, piles, and excavation levels\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left column, below section header\",\n            \"image_style\": \"none\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Technical diagram showing four vertical piles (two solid, two with rebar) penetrating silty clay and sand layers. Includes labels for ground level, excavated level, earth anchor/strut, and soil types.\"\n          }\n        },\n        {\n          \"id\": \"body-741-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SYSTEM COMPONENTS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, top\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Section header in bold, all caps, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"body-741-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• TOUCHING PILES: Closely spaced bored piles.\\n• PRIMARY FUNCTION: Temporary or permanent retaining wall.\\n• GROUNDWATER BEHAVIOUR: Limited seepage control.\\n• INSTALLATION STEPS: Bore, place cage, concrete.\\n• REINFORCEMENT DETAIL: Selected piles reinforced.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"minor\",\n            \"position\": \"right column, below section header\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bulleted list with bolded key terms followed by explanations.\"\n          }\n        },\n        {\n          \"id\": \"body-741-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"BENEFITS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, below features list\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Section header in bold, all caps, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"illustration-741-5\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Four advantage icons arranged in a 2x2 grid: a green leaf icon, a blue wind swirl icon, a blue water drop icon, and a brown building icon.\",\n          \"style\": {\n            \"illustration_style\": \"flat icons\",\n            \"color_scheme\": \"green, blue, brown\",\n            \"position\": \"right column, bottom\",\n            \"special_instructions\": \"Each icon is paired with text describing an advantage: LOW NOISE LEVELS, MINIMIZED SETTLEMENT, ADAPTABLE TO STRATIGRAPHY, SMALL FOOTPRINT.\"\n          }\n        },\n        {\n          \"id\": \"body-741-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"LOW NOISE LEVELS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"minor\",\n            \"position\": \"right column, bottom-left of advantages section\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Text label next to leaf icon, all caps.\"\n          }\n        },\n        {\n          \"id\": \"body-741-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"MINIMIZED SETTLEMENT\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"minor\",\n            \"position\": \"right column, bottom-right of advantages section\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Text label next to wind icon, all caps.\"\n          }\n        },\n        {\n          \"id\": \"body-741-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"ADAPTABLE TO STRATIGRAPHY\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"minor\",\n            \"position\": \"right column, bottom-left of advantages section, below first item\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Text label next to water icon, all caps.\"\n          }\n        },\n        {\n          \"id\": \"body-741-8\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SMALL FOOTPRINT\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"minor\",\n            \"position\": \"right column, bottom-right of advantages section, below second item\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Text label next to building icon, all caps.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nCreate this JSON as a single 16:9 slide image. Ensure the result is visually engaging, easy to read, and fully aligned with all styling requirements.", "input_images": ["images/multi_image/4a5fdbe4-f899-4648-bf0e-797174efd9bb_input_0.jpg"], "target_image": "images/multi_image/4a5fdbe4-f899-4648-bf0e-797174efd9bb_target.png", "width": 1376, "height": 768}
+{"index": 195, "id": "a027f4aa-4a4a-486b-abc8-2fcaceb1c3e3", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide features a cool, light teal background with a faint, soft-focus grid of tiny dots, giving a technical yet calm atmosphere. At the very top center, a wide, semi‑transparent white title ribbon with rounded corners displays the text **“Training Hyper‑Parameters Overview”** in a bold, dark charcoal sans‑serif font.\n\nBeneath the title, aligned centrally on the slide, the **sub image** is prominently placed. It shows a concise comparison table of hyper‑parameters for two model types, labeled **“discrete”** and **“continuous”**. The leftmost column is titled **“hyper‑parameters”**, listing:\n\n- optimizer  \n- β₁  \n- β₂  \n- weight decay  \n- lr (pre‑training)  \n- lr (fine‑tuning)  \n- lr scheduling  \n- lr warmup steps  \n- batch size  \n- training steps (pre‑training)  \n- training steps (fine‑tuning)\n\nTo the right, the table presents two vertical columns of values. Under **“discrete”**, the entries include **AdamW** as optimizer, β₁ = **0.9**, β₂ = **0.96**, weight decay **0.03**, learning rates **0.0004** and **0.0001**, **cosine** scheduling, **10K** warmup steps, batch size **4096**, and training steps **500K** and **250K**. Under **“continuous”**, the corresponding values are **AdamW**, β₁ = **0.9**, β₂ = **0.95**, weight decay **0.02**, learning rates **0.0002** and **0.0002**, **constant** scheduling, **50K** warmup steps, batch size **2048**, and training steps **1000K** and **500K**. Thin horizontal and vertical rules outline the header and separate the three main columns, emphasizing the comparison.\n\nTo keep the **sub image** as the primary focus, only two small supporting text panels appear, symmetrically placed.\n\nOn the upper left, a rounded rectangle with a faint teal border contains a short caption in a clean, dark grey font:  \n**“Side‑by‑side settings for discrete vs. continuous training regimes.”**\n\nOn the upper right, a matching rounded rectangle carries a second concise note:  \n**“Key differences: schedule type, warmup length, batch size, and total steps.”**\n\nBoth text boxes are modest in size and lightly transparent so they do not distract from the central table. The rest of the slide remains open and uncluttered, with generous margins around the **sub image**, emphasizing that the table is the main informational element for the audience.", "input_images": ["images/multi_image/a027f4aa-4a4a-486b-abc8-2fcaceb1c3e3_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 196, "id": "e09259f3-c1b9-4d84-a4c9-42125066991f", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Split between a concise textual block on the left and the sub image on the right, this presentation slide is set against a clean, off‑white background with subtle, concentric gray line patterns in the upper-left and bottom-right corners.  \n\nThe slide’s primary heading, positioned at the top left in a bold, dark charcoal sans‑serif font, reads:  \n“EVALUATING INSTANCE-CONTROLLED DIFFUSION PERFORMANCE”.  \n\nDirectly beneath this title is a compact block of smaller, black sans‑serif text stating:  \n“Comparison of adapter-based and training-free methods on instance success and segmentation quality across difficulty levels.”  \n\nBelow this description, aligned to the left margin, two short bullet points summarize the key takeaway:  \n\n• “3DIS variants consistently outperform prior training-free approaches.”  \n• “MIGC leads among adapter rendering baselines on both metrics.”  \n\nOn the right half of the slide, occupying most of the vertical space and centered both horizontally and vertically, is the sub image. The table is displayed without cropping, maintaining its original proportions so that all method names, metric columns (Instance Success Ratio and Mean Intersection over Union), and level annotations (L2–L6 and AVG) remain fully legible. Thin, soft-shadow edging around the sub image gently lifts it from the background, emphasizing it as the main focal element.  \n\nThe overall layout preserves a clear visual hierarchy: the succinct explanatory text on the left frames the content, while the sub image on the right serves as the dominant visual centerpiece illustrating quantitative performance comparisons between diffusion-based rendering methods.", "input_images": ["images/multi_image/e09259f3-c1b9-4d84-a4c9-42125066991f_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 197, "id": "81851123-6528-417b-b634-65dca3e4f287", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Laid out with a clean, educational design, this slide explores how the body responds to tissue injury. The title \"WHAT HAPPENS DURING INFLAMMATION?\" is centered at the top in a large, sans-serif font. To the left, four paragraphs of text describe the sequence of events in an inflammatory response: \"When tissues are damaged, local cells release chemical signals that call immune cells to the site.\" The text continues, \"These mediators cause nearby blood vessels to widen and become leaky, allowing plasma and white blood cells to enter the affected area.\" It further explains, \"Signals from immune cells help clear debris and invaders, but if the response is excessive or prolonged, it can injure healthy tissue,\" and concludes with, \"Recognizing how these processes unfold helps clinicians tailor therapies that control inflammation without blocking necessary healing.\" On the right side, the slide features the diagram from the first input image, which depicts a shield with an alarm bell in the center, surrounded by icons representing various triggers like \"Physical,\" \"Chemical,\" \"Infectious,\" and \"Immunological,\" as well as symptoms such as \"redness,\" \"swelling,\" \"heat,\" and \"loss of function.\"", "input_images": ["images/multi_image/81851123-6528-417b-b634-65dca3e4f287_input_0.jpg"], "target_image": "images/multi_image/81851123-6528-417b-b634-65dca3e4f287_target.png", "width": 1376, "height": 768}
+{"index": 198, "id": "4d6aa076-3345-4d74-a669-3b4e0df6d974", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This slide presents a \"COMMUNITY STRATEGY WORKSHOP\" agenda on a white background, framed by a geometric banner of blue, teal, and orange shapes. The layout is divided into three vertical columns, each headed by a specific icon: the first image showing a group and megaphone is at the top left, the second image with a podium and trees is in the center, and the third image depicting people holding signs is at the bottom right.  \n\nThe left column is titled \"INTRODUCTION & OBJECTIVES\" and lists \"Session Goals and Expectations\" and \"Facilitator Welcome.\" The center column, labeled \"KEY STRATEGIC THEMES,\" details \"Vision and Priorities,\" \"Initiative Brainstorming,\" and \"Resource Planning.\" The right column, titled \"INTERACTIVE FEEDBACK SESSION,\" outlines a \"Participant Ideas Roundtable\" and \"Identifying Challenges and Solutions.\" A footer spans the bottom with the text \"BUILDING OUR FUTURE TOGETHER | DATE | VENUE (OPTIONAL).\"", "input_images": ["images/multi_image/4d6aa076-3345-4d74-a669-3b4e0df6d974_input_0.jpg", "images/multi_image/4d6aa076-3345-4d74-a669-3b4e0df6d974_input_1.jpg", "images/multi_image/4d6aa076-3345-4d74-a669-3b4e0df6d974_input_2.jpg", "images/multi_image/4d6aa076-3345-4d74-a669-3b4e0df6d974_input_3.jpg"], "target_image": "images/multi_image/4d6aa076-3345-4d74-a669-3b4e0df6d974_target.png", "width": 1376, "height": 768}
+{"index": 199, "id": "5a8177fe-d1c2-4213-99d2-8a9f6b3954aa", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "The slide, titled in large bold capital letters as \"STRATEGIC ALIGNMENT ROADMAP,\" is structured as a flow chart moving from left to right, incorporating elements from the input images. The process begins on the left with \"VISION & DIRECTION,\" illustrated by the first sub-image showing people with a compass and the text \"FUTURE GOALS.\" This section outlines \"Clarifying Long-Term Ambitions,\" \"Establishing a Clear North Star,\" and \"Translating Mission into Targets.\" The flow continues to the middle section labeled \"COLLABORATION & PRACTICES,\" which utilizes the second and third sub-images depicting a meeting and icons for \"ETHICS\" and \"TRUST.\" This area focuses on \"Building Collaborative Routines,\" \"Embedding Integrity in Daily Work,\" and \"Strengthening Relationships & Accountability.\" The third column, \"EXECUTION & RESULTS,\" includes graphics of a clipboard, ladder, and medals, describing \"Aligning Projects with Priorities,\" \"Rewarding Impactful Contributions,\" and \"Advancing Skills & Career Pathways.\" A final large arrow at the bottom points to \"IMPACT & VALUE,\" represented by the fourth sub-image with a lightbulb and charts. This section lists benefits such as \"Higher Innovation & Creativity,\" \"Improved Operational Efficiency,\" \"Resilient High-Performing Teams,\" \"Scalable Business Expansion,\" and \"Elevated Stakeholder Satisfaction.\" The slide concludes with a copyright notice for \"Business Concepts\" in the bottom right corner.", "input_images": ["images/multi_image/5a8177fe-d1c2-4213-99d2-8a9f6b3954aa_input_0.jpg", "images/multi_image/5a8177fe-d1c2-4213-99d2-8a9f6b3954aa_input_1.jpg", "images/multi_image/5a8177fe-d1c2-4213-99d2-8a9f6b3954aa_input_2.jpg", "images/multi_image/5a8177fe-d1c2-4213-99d2-8a9f6b3954aa_input_3.jpg"], "target_image": "images/multi_image/5a8177fe-d1c2-4213-99d2-8a9f6b3954aa_target.png", "width": 1376, "height": 768}
+{"index": 200, "id": "479c28ea-e638-4f86-a0ae-c01823058745", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 4. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Neon City Orbits\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#020617\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at center, #020617 40%, #000000 100%)\\\",\\n      \\\"texture\\\": \\\"noise\\\",\\n      \\\"pattern\\\": \\\"subtle concentric circles fading out from the center in very dark blue\\\"\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Montserrat\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#38BDF8\\\",\\n      \\\"secondary\\\": \\\"#F97316\\\",\\n      \\\"text\\\": \\\"#E5E7EB\\\",\\n      \\\"accent\\\": \\\"#22C55E\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall atmosphere is futuristic and cinematic, inspired by long-exposure light trails and aerial cityscapes. Use a dark night-sky backdrop with a faint radial glow behind the central image to make the circular interchange feel like a glowing portal. Typography should be clean and geometric with subtle neon highlights. Avoid heavy borders; rely on soft outer glows and gentle drop shadows to lift elements from the dark background.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Futuristic, Dynamic, Sophisticated\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 9,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-9\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Orbiting Through the Night\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#E5E7EB\\\",\\n            \\\"font_family\\\": \\\"Montserrat\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, slightly above the image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Uppercase lettering with expanded tracking; soft cyan glow beneath the text to echo the light trails.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/479c28ea-e638-4f86-a0ae-c01823058745_input_0.png\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/479c28ea-e638-4f86-a0ae-c01823058745_input_0.png\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Aerial night view of a circular highway interchange with colorful light trails.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center of slide, spanning about 80% of width and 75% of height\\\",\\n            \\\"image_style\\\": \\\"edge-to-edge hero with subtle outer glow\\\",\\n            \\\"container\\\": \\\"rounded rectangle frame with thin cyan border\\\",\\n            \\\"special_instructions\\\": \\\"Use large rounded corners (24px radius) and a soft blue outer shadow (#38BDF8 at 40% opacity) to make the image feel like a luminous window into the city. Ensure it remains the most dominant visual element.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-9-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"A bird’s‑eye glimpse of urban motion, captured in flowing neon arcs.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#9CA3AF\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image frame\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Small, centered text with light weight; keep spacing generous so it feels airy against the dark background.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-9-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two thin concentric neon rings partially peeking from behind the main image, aligned with its circular interchange.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"behind the central image, centered on the slide\\\",\\n            \\\"special_instructions\\\": \\\"Low-opacity rings in cyan and soft orange (#38BDF8 and #F97316), slightly larger than the image, with a 30% blur to create a halo effect that reinforces the sense of motion.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nNow render the final slide as a polished 16:9 presentation image.", "input_images": ["images/multi_image/479c28ea-e638-4f86-a0ae-c01823058745_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 201, "id": "cab232f7-434d-4e7b-aea0-76689adb6502", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Set against a dark green background, this presentation slide features yellow text on the left and the sub image on the right. The slide is titled \"Tuning Models: Measuring the Gains\" in large font at the top. Below the title, two short lines of text read, \"Fine‑tuning reshapes error and accuracy across model sizes.\" and \"Comparing EER and F1 reveals the impact of KUQ training.\" To the right, the slide prominently displays the sub image—a white rectangular table contrasting original open‑source models with their (KUQ) fine‑tuned counterparts, highlighting improvements in EER and F1 scores.", "input_images": ["images/multi_image/cab232f7-434d-4e7b-aea0-76689adb6502_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 202, "id": "b0a6ed66-0e15-48d1-ac4d-5743377a3e21", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Styled with a soft, abstract background blending pastel hues of light blue, peach, and white, this presentation slide centers on educational content regarding art criticism methodologies. At the top, a large, rounded rectangular banner with a light blue background features the slide title in a dark navy blue sans-serif font, reading \"Guided Art Critique: Turning Observation into Insight.\" Below this header, the layout is divided into two distinct vertical sections. On the left side, two rounded rectangular text boxes with thin dark outlines contain instructional text in a dark sans-serif font; the upper box reads, \"Use a clear, step‑by‑step critique sequence to engage with any artwork—visual pieces, live events, or mixed-media projects,\" and the lower box states, \"Scaffolded question sets and note templates support each phase, helping learners justify their ideas with visual evidence.\" A small orange circular bullet point floats to the left of the first text box. On the right side of the slide, a sub-image depicts a diverse group of students gathered around a framed painting of a seated deity, each holding a worksheet labeled with a specific step of the critique process—visible labels include \"Description,\" \"Analysis,\" \"Interpretation,\" and \"Judgment.\" The overall aesthetic is clean and academic, accented by small decorative star and dot shapes in the corners of the background, which add a subtle touch of whimsy to the professional layout.", "input_images": ["images/multi_image/b0a6ed66-0e15-48d1-ac4d-5743377a3e21_input_0.jpg"], "target_image": "images/multi_image/b0a6ed66-0e15-48d1-ac4d-5743377a3e21_target.png", "width": 1376, "height": 768}
+{"index": 203, "id": "2c0bc29b-5eb2-4582-aeec-7163b0c0103b", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Framed in a clean, academic presentation style, this slide highlights comparative alignment error rates from recent research. The composition features the sub image on the left, showing the central results table, while concise annotations are arranged to the right. The slide header reads “ALIGNMENT METHODS: AER COMPARISON” in bold, dark lettering across the top. \n\nOn the right, four short callouts appear in a vertical column:\n\n1. **Attention-Based Approaches**  \n   Key variants compared: prior work vs. “Ours (HI),” “Ours (Mask),” and “Ours (HI + Mask).”\n\n2. **Performance Metrics**  \n   Columns report AER and standard deviation for source–target alignments \\(A_{i,j}\\) and \\(A_{t,j}\\).\n\n3. **Baselines & Alternatives**  \n   Includes vector-norm methods and classic word aligners (Fast-Align, GIZA++).\n\n4. **Observation**  \n   Lower AER values indicate stronger alignment quality across methods.\n\nAt the bottom, a slim footer bar spans the width of the slide with the caption:  \n“FOCUS: COMPARING ALIGNMENT ERROR RATES ACROSS METHODS.”", "input_images": ["images/multi_image/2c0bc29b-5eb2-4582-aeec-7163b0c0103b_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 204, "id": "caa5a2f7-a833-4f26-b2ce-f957a6383ce2", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Set against a deep, chocolate-brown gradient backdrop that darkens toward the edges, the slide presents a unified visual story in gold and cream hues. At the top center, a new headline reads: “Crafting Customer Value: The Power of Unified Strategy,” set in a bold, sans-serif metallic-gold typeface. The capitalization, punctuation, and spacing should match this phrase exactly, with no added embellishments. Beneath the headline, two supporting statements appear in a slightly smaller, cream-colored serif font, left-aligned and maintaining consistent styling. The first line reads: “Memorable brands are built when every touchpoint reflects a shared purpose and promise.” Directly below it, the second line states: “Integrated strategies align teams, tools, and timelines to deliver seamless customer value.” Both sentences use standard sentence case and include no italics, underlines, or extra punctuation beyond what is written here.\n\nCentered near the bottom of the slide is a circular emblem composed of a thin gold ring enclosing four distinct iconographic elements arranged horizontally. From left to right: a circular seal bearing the intertwined “LV” monogram (reminiscent of Louis Vuitton’s logo), rendered in gold outline; two stylized cardboard boxes — one larger and upright, one smaller and angled forward — both in cream with gold ribbon accents; a simplified human silhouette in cream; and a speech bubble icon, also in cream, suggesting communication or dialogue. This entire icon cluster mirrors the visual motif seen in the first sub-image, which isolates these same four components within the same circular frame, confirming its role as a key graphical element reused at scale on the main slide.\n\nThe composition preserves a clear visual hierarchy, guiding the viewer from the prominent headline down through the clarifying text to the emblematic icons below. The warm browns, glowing golds, and soft cream tones evoke sophistication, reliability, and integration, reinforcing the theme of unified strategy and customer value creation. There are no data charts, tables, or intricate diagrams. Decorative features remain understated yet purposeful: the gradient background adds depth without clutter, and the gold circular border around the icons neatly frames the central metaphor. Overall, the slide acts as a conceptual centerpiece, using typography, palette, and iconography together to express how coordinated, cross-channel efforts can transform scattered activities into a cohesive, value-driven customer experience.", "input_images": ["images/multi_image/caa5a2f7-a833-4f26-b2ce-f957a6383ce2_input_0.jpg"], "target_image": "images/multi_image/caa5a2f7-a833-4f26-b2ce-f957a6383ce2_target.png", "width": 1376, "height": 768}
+{"index": 205, "id": "0456af90-02c5-42f2-ac95-5de5743fbb64", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The presentation slide adopts a modern, professional look on a light gray background, using a restrained palette of black, white, and multiple shades of green. Across the very top, a strong headline in large, sans-serif capital letters declares \"OUR EXPERTISE: STRATEGIC CAPABILITIES\". The segment \"OUR EXPERTISE:\" appears in black, while \"STRATEGIC CAPABILITIES\" is highlighted in vivid lime green to draw attention to the key message. A slim light gray horizontal rule runs beneath the title, clearly separating it from the body content.\n\nCentered under the header are two white, rounded-rectangle text panels with subtle drop shadows, aligned side-by-side. The left panel states, \"Seasoned specialists bring complementary strengths across operations, regulation, capital markets, and sustainability.\" The right panel reads, \"Integrated capabilities drive disciplined decision‑making and long‑term value creation.\" Both passages are in a clean black sans-serif typeface and are left-aligned within their respective boxes.\n\nBelow these panels, a horizontal row of four square feature cards highlights the core capability areas. Each card is white with rounded corners and a gentle shadow, containing a green icon paired with a bold black label in capital letters. The first card on the left features the sub-image of a mining helmet, pickaxe, and ore pile, labeled \"OPERATIONS\" beneath it. The second card displays the sub-image of a gavel and scales of justice, labeled \"REGULATION\". The third card shows the sub-image of ascending coins with a dollar sign and a trend line, labeled \"CAPITAL\". The final card on the right presents the sub-image of a globe, leaves, and recycling symbols, labeled \"SUSTAINABILITY\". The icons across all four cards share a unified green color treatment, reinforcing visual cohesion and the slide’s polished corporate style.", "input_images": ["images/multi_image/0456af90-02c5-42f2-ac95-5de5743fbb64_input_0.jpg", "images/multi_image/0456af90-02c5-42f2-ac95-5de5743fbb64_input_1.jpg", "images/multi_image/0456af90-02c5-42f2-ac95-5de5743fbb64_input_2.jpg", "images/multi_image/0456af90-02c5-42f2-ac95-5de5743fbb64_input_3.jpg"], "target_image": "images/multi_image/0456af90-02c5-42f2-ac95-5de5743fbb64_target.png", "width": 1376, "height": 768}
+{"index": 206, "id": "033d0067-d64b-4cb5-9f71-ff1659c6d518", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### CREATIVE RENDERING BRIEF\nAct as a world-class presentation designer generating a single slide from structured JSON.\n\n### JSON SEMANTICS\n- `global_style`: the fixed visual language for this slide deck. Apply it uniformly and precisely.\n- `visual_hierarchy`: the intended viewing order, emphasis levels, and overall mood.\n- `source_images`: the images that must be placed on the slide.\n- `blocks`: the exclusive set of elements that may appear on the slide.\n- `special_instructions`: explicit art direction that overrides other guidance where specified.\n\n### NON-NEGOTIABLES\n- Do not invent or alter any text, figures, legends, icons, logos, or headings.\n- Ensure all text remains fully visible within the canvas; tune line breaks, size, spacing, and weight for clarity.\n- Maintain all referenced colors and fonts exactly. When a block cites a palette token, use that precise value.\n- If the provided content is sparse, embrace negative space. Do not introduce additional graphics or wording.\n\n### COMPOSITION GUIDANCE\nCraft a clear, period-inspired layout appropriate to the content volume. Use columns or card-like groupings for text-dense slides; let imagery dominate when visuals are primary; for mixed content, balance graphics and copy with a strong hierarchy. Output a polished, production-ready slide image.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Vintage Savings Breakdown\",\n    \"background\": {\n      \"color\": \"#F5EBCF\",\n      \"gradient\": null,\n      \"texture\": \"paper\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Cooper Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#9C7A3D\",\n      \"secondary\": \"#C0C0C0\",\n      \"text\": \"#4A1E1E\",\n      \"accent\": \"#D4A017\"\n    },\n    \"special_instructions\": \"The slide keeps a nostalgic, tactile appearance with a cream-toned, paper-like background. Use a subdued, earthy color mix led by warm browns, golds, and soft greys. Apply a light distressed or grainy finish to icons and frames. Organize the layout with a decorative border and a clearly defined header band separated from the main content.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Reassuring, Vintage, Data-Driven\",\n    \"composition_style\": \"Centered, Grid-based\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 434,\n      \"blocks\": [\n        {\n          \"id\": \"illustration-434-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative rectangular border frame with rounded corners and small inward-pointing corner accents, enclosing the entire slide content.\",\n          \"style\": {\n            \"illustration_style\": \"decorative frame\",\n            \"color_scheme\": \"#9C7A3D\",\n            \"position\": \"outermost edge of slide\",\n            \"special_instructions\": \"Thin line weight, defines the content area.\"\n          }\n        },\n        {\n          \"id\": \"illustration-434-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"An illustration of two stylized hands, one above the other, with a grainy texture, colored in gold/brown tones.\",\n          \"style\": {\n            \"illustration_style\": \"textured icon\",\n            \"color_scheme\": \"#9C7A3D and #D4A017\",\n            \"position\": \"top-left corner\",\n            \"special_instructions\": \"Part of the header decoration.\"\n          }\n        },\n        {\n          \"id\": \"title-434\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Household Savings Mix\",\n          \"style\": {\n            \"color\": \"#9C7A3D\",\n            \"font_family\": \"Cooper Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, bold serif title with a classic vintage character.\"\n          }\n        },\n        {\n          \"id\": \"body-434-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Share of Monthly Income by Category\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#4A1E1E\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center, below title\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Subtitle in a smaller, clean sans-serif style.\"\n          }\n        },\n        {\n          \"id\": \"illustration-434-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"An illustration of a balance scale. The left pan holds gold bars, and the right pan holds a dollar sign symbol.\",\n          \"style\": {\n            \"illustration_style\": \"textured icon\",\n            \"color_scheme\": \"#9C7A3D, #D4A017, #C0C0C0\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"Part of the header decoration, matches the hands icon style.\"\n          }\n        },\n        {\n          \"id\": \"illustration-434-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A thin horizontal line separating the header section from the main content area.\",\n          \"style\": {\n            \"illustration_style\": \"divider line\",\n            \"color_scheme\": \"#9C7A3D\",\n            \"position\": \"below header, spanning full width\",\n            \"special_instructions\": \"Simple separator.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_561_411_812_578.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_561_411_812_578.jpg\"\n              }\n            ],\n            \"source\": \"Circular chart showing 15%\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-left, first in row of three\",\n            \"image_style\": \"none\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Circular progress indicator with a grey ring and a gold segment.\"\n          }\n        },\n        {\n          \"id\": \"body-434-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"RAINY-DAY FUND\\n(short-term cash buffer)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#4A1E1E\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-left, below first chart\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Label for the first chart, centered under the circle.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_1011_414_1259_578.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_1011_414_1259_578.jpg\"\n              }\n            ],\n            \"source\": \"Circular chart showing 7%\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-right, third in row of three\",\n            \"image_style\": \"none\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Circular progress indicator with a grey ring and a gold segment.\"\n          }\n        },\n        {\n          \"id\": \"body-434-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"LONG-TERM INVESTMENTS\\n(retirement and growth)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#4A1E1E\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right, below third chart\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Label for the third chart, centered beneath the circle.\"\n          }\n        },\n        {\n          \"id\": \"illustration-434-5\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A circular chart graphic in the center column showing '5%' with a grey ring and a small gold segment at the top.\",\n          \"style\": {\n            \"illustration_style\": \"infographic element\",\n            \"color_scheme\": \"#C0C0C0, #D4A017, #4A1E1E\",\n            \"position\": \"bottom-center, second in row of three\",\n            \"special_instructions\": \"Matches the style of the other two charts.\"\n          }\n        },\n        {\n          \"id\": \"body-434-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"EDUCATION SAVINGS\\n(future tuition and courses)\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#4A1E1E\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-center, below middle chart\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Centered label for the middle chart.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nProduce a professional 16:9 slide image that faithfully reflects the structured JSON and avoids any invented content.", "input_images": ["images/multi_image/033d0067-d64b-4cb5-9f71-ff1659c6d518_input_0.jpg", "images/multi_image/033d0067-d64b-4cb5-9f71-ff1659c6d518_input_1.jpg"], "target_image": "images/multi_image/033d0067-d64b-4cb5-9f71-ff1659c6d518_target.png", "width": 1376, "height": 768}
+{"index": 207, "id": "d126a879-72b2-4681-9b67-95cc8d4ab9c8", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Build a slide with a soft olive‑green gradient background that darkens slightly toward the bottom, creating a professional and calming atmosphere. At the very top, centered, place a wide, pale mint‑green rectangular banner with rounded corners containing the title text. The text reads **\"Iterative AI System & Test Co‑Development\"** in a large, black, serif typeface with standard weight and no italics.\n\nBelow the title, the layout again splits horizontally into two distinct content areas. On the left, occupying roughly the lower‑left quadrant, position the rounded‑corner version of the sub image. Scale the sub image so it is clearly legible but still leaves sufficient breathing room around it. The sub image shows two vertical process flows—“Test Development” in soft pink on the left and “System Development” in light blue on the right—connected through a central bar labeled “Test system and report results” and a final step leading to deployment and observation. Ensure the full diagram, including all step labels and directional arrows, is visible and unobstructed.\n\nTo the right of this sub image, place a large, white rectangular text box with rounded corners that contains the body of the slide. The text inside is black, using a clean sans‑serif font. Keep line spacing relaxed for readability and maintain generous margins inside the box.\n\nThe first brief paragraph begins immediately with the sentence:  \n**\"Reliable AI emerges from parallel development of tests and systems, each informing the other through iterative feedback loops.\"**\n\nFollowing a small vertical gap, add a second concise paragraph:  \n**\"The diagram highlights how reliability goals, test construction, data collection, and system design converge before deployment, ensuring only tested capabilities reach real users.\"**\n\nAt the bottom of the white box, separated by another small gap, include a single line of smaller text that reads:  \n**\"Use this co‑development cycle as a template for structuring your own AI evaluation workflow.\"**\n\nNo additional decorative elements or icons are added; the composition is framed solely by the olive‑green gradient background, with the title banner at the top, the sub image on the left, and the minimal explanatory text on the right, allowing the diagram to remain the primary visual focus of the slide.", "input_images": ["images/multi_image/d126a879-72b2-4681-9b67-95cc8d4ab9c8_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 208, "id": "9b807c68-367a-44a5-8d07-3a6d3a73e0b5", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "Render a single presentation slide by interpreting the provided Structured Content JSON as your complete specification.\n\nFollow this precedence order:\n1. Treat `global_style` as the master guide for the slide’s overall visual language.\n2. Use `visual_hierarchy` to determine emphasis, layout balance, and compositional rhythm.\n3. Use `blocks` to place all text and visual elements exactly as defined.\n4. Apply `special_instructions` as high-priority creative direction that can override default styling when needed.\n5. Use all `source_images` as embedded visual assets in the slide.\n\nDo not invent, remove, or alter any textual content from the JSON. You may refine only the typography and layout: adjust line breaks, alignment, font size, weight, spacing, and hierarchy to enhance clarity and aesthetics.\n\nDesign the slide as a custom, well-composed layout rather than a generic template. Avoid repetitive bullet lists and boxy grids. Use generous whitespace, clear grouping, and precise alignment to create an organized composition that supports the message.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Medical Infographic\",\n    \"background\": {\n      \"color\": \"#D4D4F7\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": \"Geometric shapes in corners (triangles, circles, rounded rectangles) in varying colors\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2D2D2D\",\n      \"secondary\": \"#FFFFFF\",\n      \"text\": \"#2D2D2D\",\n      \"accent\": \"#D4D4F7\"\n    },\n    \"special_instructions\": \"The slide features a clean, modern medical presentation style with a soft lavender background. Decorative geometric shapes in blue, pink, and teal are scattered in the corners. The main content is contained within a large white rounded rectangle. The central visual is a stylized infographic showing a golden oil drop transforming into food items and then branching into sensory icons.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Educational, Clean\",\n    \"composition_style\": \"Centered, Vertical Stack\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 11,\n      \"blocks\": [\n        {\n          \"id\": \"title-11\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"INTEGRATED NOURISHMENT: PATHWAYS OF THERAPEUTIC OLEATION\",\n          \"style\": {\n            \"color\": \"#2D2D2D\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Set in large, bold, uppercase lettering and centered at the upper portion of the slide. Break into two lines for strong visual impact and balance above the main infographic.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_140_334_1267_662.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_140_334_1267_662.jpg\"\n              }\n            ],\n            \"source\": \"Infographic illustrating the progression of Accha Sneha into dietary elements and subsequent sensory routes\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center, occupying most of the slide width beneath the title\",\n            \"image_style\": \"displayed within a prominent white rounded rectangle container\",\n            \"container\": \"white rounded rectangle with a subtle drop shadow to lift it from the lavender background\",\n            \"special_instructions\": \"The image shows a golden droplet labeled 'Accha Sneha' at the far left, flowing into a midsection filled with food imagery (rice, milk, vegetables), then diverging on the right into golden streams that point toward minimal line-art icons representing a hand, eye, ear, nose, and medicine vials.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\nRender this JSON as a single 16:9 slide image, ensuring it is visually engaging, legible, and fully aligned with all styling directives.", "input_images": ["images/multi_image/9b807c68-367a-44a5-8d07-3a6d3a73e0b5_input_0.jpg"], "target_image": "images/multi_image/9b807c68-367a-44a5-8d07-3a6d3a73e0b5_target.png", "width": 1376, "height": 768}
+{"index": 209, "id": "b58ed63c-ea85-40d5-8d40-9d7b70d7cab5", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The slide is a clean, informational presentation titled \"Blue: Inspiring Calm and Clarity,\" designed with a white background and accented by geometric shapes in shades of blue. The layout features a main title at the top, two paragraphs of body text on the left, and a detailed data table on the right, with a summary sentence and decorative elements at the bottom.\n\nAt the top center, the title \"Blue: Inspiring Calm and Clarity\" is written in a large, bold, dark blue sans-serif font. To the right of the title, there is a solid dark blue circle and a vertical arrangement of four small, dark blue triangles pointing upwards. A larger, light blue circular shape with a wave-like edge partially frames the title area on the right side.\n\nOn the left side of the slide, beneath the title, are two paragraphs of text in a smaller, black sans-serif font. The first paragraph reads: \"Blue conveys stability and peace—ideal for spaces that soothe and reassure.\" The second paragraph states: \"Its association with trust makes blue a favorite for brands and professional settings.\"\n\nOn the right side of the slide, occupying the majority of the space, is a data table enclosed within a light blue rounded rectangle border. The table itself has a yellow header row and a light beige body. It is organized into seven columns and two rows of data. The column headers, written in white sans-serif font against the yellow background, are: \"Color Name\", \"Personality Trait\", \"Meaning\", \"Representation\", \"Psychological Effect\", \"Cultural Context\", and \"Religious Context\".\n\nThe data within the table is presented as follows:\n- **Row 1**:\n  - Under \"Color Name\": \"Blue\" (black, sans-serif).\n  - Under \"Personality Trait\": This cell is empty in the first row but is filled in the second row (see below).\n  - Under \"Meaning\": \"Calm, Trust, Reliability\" (black, sans-serif, stacked vertically).\n  - Under \"Representation\": \"Sky, Ocean, Professionalism\" (black, sans-serif, stacked vertically).\n  - Under \"Psychological Effect\": \"Promotes relaxation and focus but can feel cold or distant in excess\" (black, sans-serif, stacked vertically).\n  - Under \"Cultural Context\": \"Symbol of harmony in many cultures; associated with authority in uniforms\" (black, sans-serif, stacked vertically).\n  - Under \"Religious Context\": \"Linked with devotion and protection in various religious traditions\" (black, sans-serif, stacked vertically).\n- **Row 2**:\n  - The first cell is empty.\n  - Under \"Personality Trait\": \"Calm, Dependable, Thoughtful\" (black, sans-serif, stacked vertically).\n  - The remaining cells in this row are empty, effectively merging the content visually with the row above, though the grid lines remain visible.\n\nBelow the table, a summary sentence is centered horizontally in a dark blue sans-serif font: \"Blue’s effects: serenity, trust, focus, emotional distance.\"\n\nDecorative elements include a large, light blue rounded rectangle in the bottom-left corner containing four small, dark blue triangles pointing upwards. A solid dark blue circle sits near the bottom center. In the bottom-right corner, there is a stack of three horizontal light blue lines resembling a menu icon.", "input_images": ["images/multi_image/b58ed63c-ea85-40d5-8d40-9d7b70d7cab5_input_0.jpg"], "target_image": "images/multi_image/b58ed63c-ea85-40d5-8d40-9d7b70d7cab5_target.png", "width": 1376, "height": 768}
+{"index": 210, "id": "a6795172-987a-47fb-a930-4e2c02bef5a6", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, titled \"How Does Message Passing Work?\" in large yellow text, features a two-column layout. On the left, a sub-image displays three cartoon animals—two bears and a rabbit—looking at photos which merge into a glowing orb above them, accompanied by a blue text box explaining, \"Each of us studies our own picture first, then we send summaries to the group! The most helpful clues from every picture get fused together.\" Above this, a red banner states, \"It’s like a team brainstorm where everyone shares notes, and the final idea comes from all of us.\" The right side of the slide contains a technical diagram of a neural network architecture labeled \"Encoder\" and \"Decoder,\" showing components like \"max-pool,\" \"concat,\" \"conv,\" \"deconv,\" and \"skip connection.\" Below this diagram, a specific formula is highlighted: \"$h_i^{agg} = \\frac{1}{K} \\sum_k h_i^k$,\" with the note \"where $h_i^k$ is the $i$-th feature from the $k$-th input, and $h_i^{agg}$ is the aggregated feature across all inputs.\" The slide concludes with a caption at the bottom: \"Zoomed-in illustration of encoder and decoder blocks using skip connections and pooling layers to share information between parallel streams.\"", "input_images": ["images/multi_image/a6795172-987a-47fb-a930-4e2c02bef5a6_input_0.jpg", "images/multi_image/a6795172-987a-47fb-a930-4e2c02bef5a6_input_1.jpg"], "target_image": "images/multi_image/a6795172-987a-47fb-a930-4e2c02bef5a6_target.png", "width": 1376, "height": 768}
+{"index": 211, "id": "52e9bab6-9d8b-418f-8404-06c4c89934ab", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The presentation slide, set against a clean white background, is titled \"DECISION-MAKING STYLES (WITH RICK AND MORTY)\" in large, bold, black sans-serif text centered at the top. The layout is a comparative diagram split into two distinct conceptual zones, connected by a central horizontal arrow labeled \"THE CHOICES SPECTRUM\" in bold black capital letters, flanked by arrowheads pointing left and right.\n\nOn the left side, under the subheading \"THE 'MORTY' MODE (PLAYING IT SAFE)\" in bold black text, is a depiction of the character Morty Smith. This visual element, corresponding to the input sub-image, shows Morty standing anxiously on a small wooden platform surrounded by a shallow puddle of water. He is clutching a grey locked money box filled with stacks of green cash. Floating to his left is a column of black text: \"PRIORITIZES SECURITY OVER PROGRESS\" positioned above a pink piggy bank icon, followed by \"DODGES ANY POSSIBILITY OF LOSS\" above a green dollar bill icon, and finally \"LETS VALUABLE CHANCES SLIP AWAY.\" Floating around Morty are two illustrations of documents labeled \"SAVINGS ACCOUNT\" in black text on a light blue background.\n\nOn the right side, labeled \"THE 'RICK' MODE (BOLD EXPERIMENTS)\" in bold black text, is a dynamic illustration of the character Rick Sanchez running toward the right edge of the frame. Rick is depicted with wild blue hair, wearing a white lab coat, and holding a complex scientific portal gun emitting a blue swirling portal and green fluid. He is leaping over a jagged, dark grey cliff edge. Near his feet, a mathematical formula is displayed in black text: \"x = (t-t₀)² e⁻ᵍᶻL² / f(w) = +∫ z⁻ᵍᵉᵍᶻᵍᶻc / ω = ∫ z (t-t₀)² + B²\". Below the cliff, black text reads \"CAN SPIRAL INTO CHAOS AND MELTDOWNS.\" Rick is reaching for a large, glowing golden treasure chest labeled \"UNLIMITED SQUANCH\" in black text. To the right of the chest, text reads \"CHASES MASSIVE UPSIDE (IGNORING BIG DOWNSIDE)\" in bold black letters. In the background, three mushroom cloud explosions are drawn in grey outlines. In the bottom right corner, enclosed in parentheses, is the text \"(OCCASIONALLY THE JACKPOT IS COLOSSAL)\" in black. The entire right side conveys a sense of high-energy danger and potential reward, contrasting with the static, cautious nature of the left side.", "input_images": ["images/multi_image/52e9bab6-9d8b-418f-8404-06c4c89934ab_input_0.jpg"], "target_image": "images/multi_image/52e9bab6-9d8b-418f-8404-06c4c89934ab_target.png", "width": 1376, "height": 768}
+{"index": 212, "id": "2d127541-85d9-45e9-94b9-5ede70146ad8", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "On a dark background with large, abstract lime-green numbers, the slide is titled \"CLOUD-NATIVE DELIVERY MODEL\" in bold, light green capital letters, with the subtitle \"Accelerating Innovation and Reliability\" below it. A list of bullet points appears on the left, reading: \"- Loosely Coupled Service Boundaries,\" \"- Distributed Data Ownership,\" \"- Heterogeneous Tech Stack Friendly,\" and \"- Fault-Tolerant & Elastic.\" In the center, a smaller text note states, \"9: Fine-grained, business-aligned capabilities.\" To the right, the sub-image shows a group of three professionals collaborating at a desk; overlaid on this photo is the text \"2: Independently deployable\" in white on a black box.", "input_images": ["images/multi_image/2d127541-85d9-45e9-94b9-5ede70146ad8_input_0.jpg"], "target_image": "images/multi_image/2d127541-85d9-45e9-94b9-5ede70146ad8_target.png", "width": 1376, "height": 768}
+{"index": 213, "id": "126e83c1-010f-4a96-8b14-b86b2d795e43", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are an AI slide-generation engine. Convert the supplied slide JSON into a high-quality presentation image.\n\nFollow these constraints exactly:\n- The JSON is authoritative. Render only the contents of `blocks`; never add new words.\n- Embed all source images into the slide.\n- `global_style` defines the shared visual identity. Preserve its fonts, colors, background, and style language.\n- `visual_hierarchy` describes what should feel dominant, secondary, or subtle.\n- `special_instructions` may refine or override the default look; respect it when present.\n\nDesign with variety. Do not default to a title plus bullet list. Use asymmetry, columns, modular cards, large focal visuals, restrained annotations, or generous negative space when appropriate. The final slide must feel clean, readable, and presentation-ready.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Midnight Over the Thames\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#030814\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top-center, #1B2741 0%, #030814 70%)\\\",\\n      \\\"texture\\\": \\\"noise\\\",\\n      \\\"pattern\\\": \\\"Subtle soft vignette around the edges to frame the central bridge image\\\"\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Playfair Display\\\",\\n      \\\"secondary\\\": \\\"Lato\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#E8C16A\\\",\\n      \\\"secondary\\\": \\\"#4DA6D9\\\",\\n      \\\"text\\\": \\\"#F5F7FA\\\",\\n      \\\"accent\\\": \\\"#FF6B4A\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall mood is calm, elegant, and nocturnal, echoing the blue evening sky and warm golden lights on the bridge. Background is a deep navy with a soft radial glow behind the image to subtly highlight it without distraction. Text and accents use warm golds and cool cyans picked from the bridge illumination and water reflections. Minimal decorative elements keep the scene cinematic and allow the illuminated architecture to dominate.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Elegant, Serene, Travel-inspired\\\",\\n    \\\"composition_style\\\": \\\"Full-bleed image with overlaid title\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 24,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/126e83c1-010f-4a96-8b14-b86b2d795e43_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/126e83c1-010f-4a96-8b14-b86b2d795e43_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Illuminated suspension bridge over a river at night with boats in the foreground.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning full width and about 80% of slide height\\\",\\n            \\\"image_style\\\": \\\"edge-to-edge with subtle soft shadow on top and bottom edges\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Maintain full color richness; apply very slight vignette on left and right edges to keep focus toward the center towers. No border radius to preserve a cinematic, panoramic feel.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"title-24\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Night Lights on the River\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#F5F7FA\\\",\\n            \\\"font_family\\\": \\\"Playfair Display\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"overlaid on image, bottom-left corner\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"banner\\\",\\n            \\\"special_instructions\\\": \\\"Title sits within a semi-transparent deep navy banner (#030814) with 70% opacity, spanning roughly 55% of the slide width and hugging the bottom-left. Text is left-aligned, with a soft drop shadow to preserve legibility over the water reflections.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-24-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"An iconic city bridge glowing against the calm twilight sky.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Lato\\\",\\n            \\\"color\\\": \\\"#E8C16A\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"directly below the title within the same bottom-left banner\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Smaller size than the title, left-aligned, with generous letter spacing for an airy, refined feel.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-24-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A thin, horizontal gold line serving as an accent underline beneath the title banner, extending slightly beyond its right edge.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"minimalist line art\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"just below the bottom-left title/banner area, overlapping slightly with the lower edge of the image\\\",\\n            \\\"special_instructions\\\": \\\"Line is about 3px thick, 60% of slide width, with a soft glow in #E8C16A fading toward the ends to subtly echo the bridge lights.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate a single widescreen slide from the JSON. The output should be refined, readable, and consistent with the specified visual system.", "input_images": ["images/multi_image/126e83c1-010f-4a96-8b14-b86b2d795e43_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 214, "id": "17954bf5-12ef-4fef-bf26-92d8caca2a63", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Against a solid yellow background, the slide features a large, black, all-caps title \"STRATEGIC STORYTELLING\" in the top-left corner. Directly below the title is the input image showing a person's tattooed hands typing on a laptop keyboard. To the right of the image, four colored text boxes are stacked vertically. The top purple box reads \"CLARIFY THE CORE MESSAGE. DEFINE A SINGLE, MEMORABLE TAKEAWAY AND ENSURE EVERY DETAIL REINFORCES THAT CENTRAL POINT.\" Below it, a green box states \"CONNECT FACTS TO IMPACT. LINK KEY METRICS TO REAL-WORLD OUTCOMES SO THE AUDIENCE UNDERSTANDS WHY THE NUMBERS MATTER.\" The third box, in blue, contains the text \"ILLUSTRATE THE JOURNEY. HIGHLIGHT WHERE YOU STARTED, PIVOTAL DECISIONS MADE ALONG THE WAY, AND THE PROGRESS ACHIEVED.\" Finally, the bottom red box reads \"END WITH A CLEAR NEXT STEP. PROVIDE A SPECIFIC INVITATION, COMMITMENT, OR CALL TO ACTION SO THE STORY LEADS TO MEANINGFUL FOLLOW-THROUGH.\"\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/17954bf5-12ef-4fef-bf26-92d8caca2a63_input_0.jpg"], "target_image": "images/multi_image/17954bf5-12ef-4fef-bf26-92d8caca2a63_target.png", "width": 1376, "height": 768}
+{"index": 215, "id": "5b7f2b34-e6ba-4d90-947d-62ba21b38cc5", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, titled \"POWER MANAGEMENT: SMART PERFORMANCE CONTROL\" in bold brown text at the top, uses cute cartoon characters to explain how a computer balances speed and efficiency. The slide layout features a top section comparing \"HIGH EFFICIENCY MODE\" on the left with \"PERFORMANCE LIMITS\" on the right, where the left panel is the second sub-image and the right panel is the fourth sub-image. Below these panels are two columns labeled \"WHY IT HAPPENS\" and \"WHAT YOU CAN DO.\" The \"WHY IT HAPPENS\" section lists several contributing conditions, incorporating the third sub-image which shows a sun and thermometer labeled \"HIGH AMBIENT TEMP\" as the third point. The \"WHAT YOU CAN DO\" section summarizes slowdowns and three specific user tips illustrated by icons: \"Use a cooling pad,\" \"Avoid blocking vents,\" \"Update system software,\" and \"Take short breaks between heavy tasks.\" The overall design is clean and colorful, utilizing distinct illustrations to guide the viewer through the technical explanation of smart performance control.", "input_images": ["images/multi_image/5b7f2b34-e6ba-4d90-947d-62ba21b38cc5_input_0.jpg", "images/multi_image/5b7f2b34-e6ba-4d90-947d-62ba21b38cc5_input_1.jpg", "images/multi_image/5b7f2b34-e6ba-4d90-947d-62ba21b38cc5_input_2.jpg", "images/multi_image/5b7f2b34-e6ba-4d90-947d-62ba21b38cc5_input_3.jpg"], "target_image": "images/multi_image/5b7f2b34-e6ba-4d90-947d-62ba21b38cc5_target.png", "width": 1376, "height": 768}
+{"index": 216, "id": "dfb58a2b-64b7-44c0-8839-174a9a84c8dd", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Design this slide to introduce the theme **\"WINTER ROOFTOP PORTRAIT\"** with the subtitle **\"LIGHT, TEXTURE, & MOOD\"** centered in bold dark blue text at the top. The layout is split into three columns: the left column highlights context and mood, the center showcases the main sub image, and the right column notes visual details.\n\nIn the center, the **sub image** (the provided photo of a person with wind‑tousled hair on a rooftop at dusk) is placed prominently, occupying most of the slide height. It is framed by a thin white border and slightly elevated from the background to draw focus.\n\nOn the left, under the bold header **\"SCENE & SETTING\"**, keep a short bullet list:\n- Rooftop at blue hour  \n- Soft, natural light  \n- Calm city backdrop  \n\nBelow this, a second mini-section titled **\"MOOD\"** includes:\n- Quiet, introspective moment  \n- Gentle winter atmosphere  \n\nOn the right side, under **\"VISUAL NOTES\"**, use concise bullets:\n- Windswept blonde hair  \n- Cozy light sweater textures  \n- Shallow depth of field  \n\nBeneath that, a final line in smaller text reads: **\"Focus on the portrait; words stay minimal.\"** At the very bottom, a narrow footer centered in muted gray states: **\"Exploring emotion through a single winter frame.\"**", "input_images": ["images/multi_image/dfb58a2b-64b7-44c0-8839-174a9a84c8dd_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 217, "id": "b27cafb6-065f-4149-ac3e-8243019073f0", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Titled **\"HAPPINESS MOMENTS\"** in large white font beneath a banner reading **\"Simple Joys by the Sea,\"** this slide features a clean two-part layout that keeps text minimal so the sub image remains the main focus.  \n\nOn the left, the sub image of a woman joyfully jumping on the beach under skywriting that reads **\"happiness is...\"** fills most of the space.  \n\nOn the right, a soft blue panel titled **\"WHAT IT FEELS LIKE\"** lists three short prompts:  \n* *Feeling free*  \n* *Living the moment*  \n* *Smiling inside and out*", "input_images": ["images/multi_image/b27cafb6-065f-4149-ac3e-8243019073f0_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 218, "id": "55e1e58a-f68f-419e-9f8f-e3f19b094fe9", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The slide features a minimalist white background with a prominent central title block and a horizontal process diagram that visualizes a progression from left to right. At the top center, the word \"PROJECT JOURNEY\" appears in large uppercase letters, styled with a playful multi-colored gradient that moves from warm reds and oranges on the left to cool greens and blues on the right. Directly below, a subtitle in a smaller gray sans-serif font reads \"From First Ideas to Finished Results.\"\n\nAcross the lower two-thirds of the slide stretches a broad, flowing arrow that runs from the left edge to the right edge, rising gently upward as it moves along. The arrow starts in a soft green hue on the left, transitions into a warm peachy orange at the midpoint, and ends in a light blue shade at the pointed tip on the right. Positioned along this arrow are four clearly defined illustration panels, each topped with a beige ribbon-style banner.\n\nThe first panel on the far left retains the banner text \"IDEATION & PLANNING.\" Below it, two children are seated at a table, working on blueprints with a sprouting plant and a ruler nearby. This corresponds to the first sub-image.\n\nShifting to the right, the second panel keeps the banner labeled \"LEARNING & SKILLS.\" The illustration here shows two children reading books, surrounded by icons of gears, lightbulbs, and question marks. This area aligns with the second sub-image.\n\nNext, the third panel maintains the title \"TEAMWORK & EFFORT.\" It depicts a group of four children collaborating to build a wooden structure, with tools and wood scraps scattered around them. This scene matches the third sub-image.\n\nThe last panel on the far right continues to be labeled \"ACHIEVEMENT & PROGRESS.\" It portrays a group of children celebrating in front of a completed windmill under a sunny sky. This corresponds to the fourth sub-image. The composition is arranged so the viewer’s eye naturally follows the rising arrow through these four stages, conveying a sense of momentum and completion.", "input_images": ["images/multi_image/55e1e58a-f68f-419e-9f8f-e3f19b094fe9_input_0.jpg"], "target_image": "images/multi_image/55e1e58a-f68f-419e-9f8f-e3f19b094fe9_target.png", "width": 1376, "height": 768}
+{"index": 219, "id": "f5daec5d-a184-4a4b-82d8-7f4e4eade269", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Design a slide titled \"RAILWAY BASICS WITH SHIN-CHAN\" using the cartoon character Shin-chan to introduce key ideas about train tracks. Keep a clear multi-section layout: on the left, place Shin-chan beside the track, with the heading \"WHAT IS A RAILWAY TRACK?\" and a short explanation: \"A railway track guides trains using two parallel rails fixed on sleepers over a bed of stones.\" Under this description, include a concise bullet list: \"Keeps trains aligned,\" \"Shares weight with the ground,\" and \"Provides a smooth path for wheels.\" Use the sub-image where Shin-chan is pointing at the rails in the upper-left portion.\n\nDirectly beneath this, show the detailed cross‑section diagram of the rails and track structure. Add a label at the top of this diagram reading \"PARTS OF THE TRACK.\" Replace the previous wording with new labels while keeping the technical terms on the drawing: a caption that says \"Rails carry the wheels,\" \"Sleepers hold the rails in position,\" \"Ballast supports drainage and stability,\" and the term \"GAUGE\" retained to mark the spacing. Do not alter any visual labels on the input cross-section image such as \"Inside rail edges,\" \"Rail head,\" \"Sleepers,\" \"Ballast,\" or \"GAUGE,\" but you may add explanatory text around it.\n\nOn the right side of the slide, feature a section titled \"DIFFERENT TRAINS, DIFFERENT TRACKS.\" Use the existing map-and-trains sub-image in the upper-right area without changing its internal labels like \"BROAD GAUGE TRAIN,\" \"STANDARD GAUGE TRAIN,\" \"NARROW GAUGE TRAIN,\" or the gauge descriptions. Next to or beneath this image, add new explanatory text such as:  \n• \"Some countries use wider tracks for high‑capacity trains.\"  \n• \"Standard gauge makes it easier for trains to run across borders.\"  \n• \"Narrow gauge is often used on mountain and regional lines.\"\n\nIn the lower-right corner, place the sub-image of Shin-chan and the woman beside the sign. Keep the sign text \"SAFETY & GAUGE\" as shown, but replace the surrounding caption with: \"Matching trains to the right track width is vital for safe, comfortable, and reliable journeys.\" Optionally, add a small note beneath: \"Engineers carefully maintain gauge so wheels stay securely on the rails.\"\n\nEnsure the overall style remains colorful and educational, matching the playful tone of the original, and arrange all four provided sub-images neatly within the slide.", "input_images": ["images/multi_image/f5daec5d-a184-4a4b-82d8-7f4e4eade269_input_0.jpg", "images/multi_image/f5daec5d-a184-4a4b-82d8-7f4e4eade269_input_1.jpg", "images/multi_image/f5daec5d-a184-4a4b-82d8-7f4e4eade269_input_2.jpg", "images/multi_image/f5daec5d-a184-4a4b-82d8-7f4e4eade269_input_3.jpg"], "target_image": "images/multi_image/f5daec5d-a184-4a4b-82d8-7f4e4eade269_target.png", "width": 1376, "height": 768}
+{"index": 220, "id": "abe99c88-b11a-4eab-bc40-8fafb8d01d40", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are an elite AI Slide Composition Engine. Your capabilities merge the sensibility of a museum exhibition designer with the rigor of an information architect.\n\n**Your Goal:** Transform the incoming **Structured Content JSON** into a single, meticulously composed presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object describing ONE slide. Interpret it as follows:\n1. **`global_style`**: Establishes the unchanging visual “universe” for the slide (background, materials, palette).\n2. **`visual_hierarchy`**: Defines layout logic, emphasis, and atmosphere.\n3. **`blocks`**: Individual elements (text, images, illustrations, shapes) that must be placed on the canvas.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INCLUSION & FIDELITY (TOP PRIORITY)\n* **Requirement:** For any block with `type: \"image\"`, you MUST place the referenced source image(s) on the slide.\n* **Source Assets:** Images are attached alongside this prompt and referenced via paths in the JSON. These exact files must appear.\n* **Protection Rules:**\n  * Do **NOT** crop, warp, recolor, repaint, or otherwise alter the pixels inside each image.\n  * Do **NOT** substitute the images with other visuals or regenerated content.\n  * Do **NOT** omit any referenced image from the final layout.\n* **Permitted Enhancements:** You MAY style the image container (frame, border, drop shadow, perspective, subtle glow) as long as the original image remains intact and fully legible.\n* **Multiple Images:** When several images are referenced, arrange them coherently (rows, columns, or according to `special_instructions`) while preserving clarity and balance.\n\n#### 2. TEXT USAGE & LIMITATIONS\n* **Exclusive Source:** All visible text must come from JSON `content` fields.\n* **Constraints:**\n  * **No invented copy:** Do not fabricate titles, labels, or filler phrases.\n  * **No verbosity padding:** Leave open space rather than creating extra wording.\n* **Typography Handling:** You may break lines, adjust alignment, or vary weight and size to enhance readability, avoiding dense, uninterrupted paragraphs.\n* **`role: \"script\"` Blocks:** Treat `script` content as background context only. Do **NOT** display it unless `special_instructions` explicitly instructs you to render it.\n\n#### 3. THE `special_instructions` PRIORITY CHANNEL\n* **Definition:** Any `special_instructions` string (in `global_style` or individual `blocks`) overrides default behavior for styling and composition.\n* **Action:** When present, execute these instructions with precedence.\n  * Example: If general rules suggest a flat layout but `special_instructions` calls for “lacquered, reflective panels and spotlighted objects,” you must follow the latter.\n* Use this field to decide on decorative texture, lighting atmosphere, and advanced effects (glow, depth of field, embossed edges, etc.).\n\n#### 4. GLOBAL_STYLE CONSISTENCY (RIGID)\n* **Obligation:** `global_style` sets the master look that must be honored consistently.\n* **Background:** Use the exact `background.color` and defined material/texture. Every slide generated under this style must share this backdrop.\n* **Font System:** Use `font_palette.primary` for all title roles and `font_palette.secondary` for supporting/body roles, with no substitutions.\n* **Color System:** When a block references a `color_ref`, use the precise hex from `color_palette`.\n* **Role-Based Sizing:**\n  * All TITLE text with `importance: \"dominant\"` must share identical size and canonical title placement.\n  * All supporting text roles (e.g., section headers, lists) must use consistent size/weight pairings across slides.\n  * Caption or minor text must also maintain a standard size and weight.\n* **Objective:** Any slides produced independently but with the same `global_style` should appear as parts of a cohesive deck—matching typography, color usage, and structural rhythm.\n\n#### 5. LAYOUT VARIETY & COMPOSITIONAL INTEREST\n* **Guideline:** Avoid defaulting to a simple “title at top + list beneath” structure.\n* **Execution Strategies:**\n  * **Text-centric layouts:** Employ columns, offset blocks, and staggered placement rather than a single vertical stack.\n  * **Image-led layouts:** When visuals dominate, allow them to occupy most of the canvas, with minimal, strategically placed text.\n  * **Hybrid layouts:** Combine visuals and copy via side-by-side groupings, overlapping frames, or layered panels.\n  * **Whitespace Discipline:** Empty areas are intentional; never stretch or inflate content to fill space.\n* **Variation Techniques:**\n  * Shift text anchoring (left/right/center) across different slides.\n  * Employ different panel shapes or clusters to group related information.\n  * Modulate the balance between heavy and light regions on the slide.\n* **Goal:** Each slide should have a distinct composition while remaining within the same visual family defined by `global_style`.\n\n### OUTPUT GENERATION\nRender a single 16:9 slide image according to these directives.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Traditional Chinese Heritage\",\n    \"background\": {\n      \"color\": \"#FDF6E9\",\n      \"gradient\": null,\n      \"texture\": \"paper\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Times New Roman\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#8B0000\",\n      \"secondary\": \"#D4AF37\",\n      \"text\": \"#3E2723\",\n      \"accent\": \"#A52A2A\"\n    },\n    \"special_instructions\": \"Use a parchment-like cream background with subtle fiber texture, framed by stylized floral and cloud motifs echoing classical Chinese paintings. Emphasize deep reds and muted golds for key accents, with refined serif title text and clean sans-serif body copy.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Cultural, Traditional, Informative, Elegant\",\n    \"composition_style\": \"Grid-based with decorative frame\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 1015,\n      \"blocks\": [\n        {\n          \"id\": \"title-1015\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"TRADITIONAL CHINESE ENGAGEMENT SEQUENCE\",\n          \"style\": {\n            \"color\": \"#3E2723\",\n            \"font_family\": \"Times New Roman\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": null,\n            \"special_instructions\": \"Prominent, centered heading with generous letter-spacing.\"\n          }\n        },\n        {\n          \"id\": \"illustration-1015-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative border elements featuring stylized peony flowers, plum blossoms, magpies, and traditional cloud motifs in the four corners and along the edges of the slide.\",\n          \"style\": {\n            \"illustration_style\": \"decorative border\",\n            \"color_scheme\": \"red, pink, blue, and gold on cream background\",\n            \"position\": \"perimeter of the slide\",\n            \"special_instructions\": \"Creates a framed effect around the content.\"\n          }\n        },\n        {\n          \"id\": \"body-1015-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"INITIAL VISIT\\n• Go-between conveys the groom’s family intentions.\\n• Families exchange birth details to review compatibility.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3E2723\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left quadrant, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Compact bullet list aligned to the left.\"\n          }\n        },\n        {\n          \"id\": \"image-1015-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_115_160_407_333.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_115_160_407_333.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a matchmaker visiting a bride's family\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-left quadrant, to the left of the Proposal & Matchmaking text\",\n            \"image_style\": \"flat illustration\",\n            \"container\": null,\n            \"special_instructions\": \"Depicts an elderly woman (matchmaker) presenting a fan to a younger woman.\"\n          }\n        },\n        {\n          \"id\": \"body-1015-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SYMBOLIC EMBLEMS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3E2723\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-right quadrant, below title\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Acts as a section header; use bold styling.\"\n          }\n        },\n        {\n          \"id\": \"illustration-1015-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A detailed illustration of a traditional Chinese dragon and phoenix facing each other.\",\n          \"style\": {\n            \"illustration_style\": \"traditional icon\",\n            \"color_scheme\": \"red, gold, and green\",\n            \"position\": \"top-right quadrant, below the Key Rituals header\",\n            \"special_instructions\": \"Symbolizes harmonious marriage.\"\n          }\n        },\n        {\n          \"id\": \"body-1015-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"DRAGON & PHOENIX: Paired motifs representing balance and conjugal harmony.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3E2723\",\n            \"importance\": \"minor\",\n            \"position\": \"top-right quadrant, to the right of the dragon/phoenix illustration\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Short descriptive line adjacent to the emblem.\"\n          }\n        },\n        {\n          \"id\": \"body-1015-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"FORMAL GIFTS (BETROTHAL SET)\\n• Groom’s side delivers trunks of textiles, jewelry, and delicacies.\\n• Demonstrates sincerity and readiness to support the new household.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3E2723\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left quadrant\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Neatly spaced bullet points, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"image-1015-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_455_330_749_522.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_455_330_749_522.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of betrothal gifts\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center of the slide, spanning horizontally\",\n            \"image_style\": \"flat illustration\",\n            \"container\": null,\n            \"special_instructions\": \"Depicts a large wooden chest filled with gifts, surrounded by family members.\"\n          }\n        },\n        {\n          \"id\": \"illustration-1015-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A collection of auspicious foods including dates, peanuts, lotus seeds, and longans.\",\n          \"style\": {\n            \"illustration_style\": \"flat icon set\",\n            \"color_scheme\": \"natural browns and reds\",\n            \"position\": \"middle-right quadrant\",\n            \"special_instructions\": \"Represents early arrival of children and prosperity.\"\n          }\n        },\n        {\n          \"id\": \"body-1015-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"AUSPICIOUS FOODS: Homophonic blessings for fertility, abundance, and sweetness in marriage.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3E2723\",\n            \"importance\": \"minor\",\n            \"position\": \"middle-right quadrant, to the right of the food illustration\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Concise explanatory caption.\"\n          }\n        },\n        {\n          \"id\": \"image-1015-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_128_529_421_689.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_128_529_421_689.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of return of gifts\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-left quadrant\",\n            \"image_style\": \"flat illustration\",\n            \"container\": null,\n            \"special_instructions\": \"Depicts a family packing gifts into boxes and a scroll.\"\n          }\n        },\n        {\n          \"id\": \"body-1015-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"RECIPROCAL GIFTS\\n• Bride’s family returns selected items from the betrothal set.\\n• Signals acceptance of the proposal and mutual courtesy.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3E2723\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-center, to the right of the return of gifts image\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Left-aligned bullets that mirror the upper sections.\"\n          }\n        },\n        {\n          \"id\": \"image-1015-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_806_511_1017_679.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_806_511_1017_679.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of double happiness symbol and red packets\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-right quadrant\",\n            \"image_style\": \"flat illustration\",\n            \"container\": null,\n            \"special_instructions\": \"Depicts a couple exchanging red packets with the double happiness symbol.\"\n          }\n        },\n        {\n          \"id\": \"body-1015-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"DOUBLE HAPPINESS & RED ENVELOPES\\n• Character for joined joy decorates boxes, banners, and packets.\\n• Monetary gifts inside envelopes convey blessings and shared fortune.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#3E2723\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right quadrant, to the right of the double happiness image\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Two-line caption clarifying the practice.\"\n          }\n        },\n        {\n          \"id\": \"illustration-1015-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative banner or ribbon element at the bottom center of the slide, featuring two white birds flying upwards.\",\n          \"style\": {\n            \"illustration_style\": \"decorative element\",\n            \"color_scheme\": \"white and gold\",\n            \"position\": \"bottom-center\",\n            \"special_instructions\": \"Adds a finishing touch to the composition.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```", "input_images": ["images/multi_image/abe99c88-b11a-4eab-bc40-8fafb8d01d40_input_0.jpg", "images/multi_image/abe99c88-b11a-4eab-bc40-8fafb8d01d40_input_1.jpg", "images/multi_image/abe99c88-b11a-4eab-bc40-8fafb8d01d40_input_2.jpg", "images/multi_image/abe99c88-b11a-4eab-bc40-8fafb8d01d40_input_3.jpg"], "target_image": "images/multi_image/abe99c88-b11a-4eab-bc40-8fafb8d01d40_target.png", "width": 1376, "height": 768}
+{"index": 221, "id": "7e5bea6e-5963-42be-b2c3-21adbd375c16", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Propaganda-Style Poster Composer. Your architecture fuses the imagination of a narrative illustrator with the discipline of a production print designer.\n\n**Your Goal:** Transform the provided **Structured Content JSON** into a single, exhibition-grade propaganda poster image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE poster. Interpret the fields as follows:\n1.  **`global_style`**: Establishes the unchanging “visual physics” of the poster (paper stock, ink character, palette, border treatment).\n2.  **`visual_hierarchy`**: Determines the arrangement logic, focal points, and emotional temperature.\n3.  **`blocks`**: The discrete elements (text, embedded images, ornaments) that must appear in the final artwork.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** Whenever a block has `type: \"image\"`, you MUST place the supplied source image files into the composition.\n*   **Input Images:** The source images are provided as attachments along with this prompt. They are the ONLY images you may use for those blocks.\n*   **Strict Constraint:** You must **PRESERVE** the pixel data of each source image.\n    *   **DO NOT** crop, blur, repaint, or otherwise alter the internal contents.\n    *   **DO NOT** swap these assets for generic illustrations or AI-generated stand‑ins.\n    *   **DO NOT** omit or hide the images—each referenced file must be visibly used.\n*   **Allowed Actions:** You MAY style the *frames* around images (e.g., add halftone edging, faux misregistration, drop shadows, or perspective skew) so long as the imagery itself remains sharp and unobstructed.\n*   **Multiple Images:** When more than one image exists, lay them out clearly (mirrored, stacked, tiled, or as directed by `special_instructions`) while maintaining balance.\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is the ONLY reservoir of copy you may render.\n*   **Strict Constraint:**\n    *   **NO Additions:** Do not invent slogans, taglines, or explanatory blurbs beyond what is provided.\n    *   **NO Padding:** Empty regions should remain intentional; use negative space as a design asset instead of filling it with new text.\n    *   **Formatting:** You may adjust size, line breaks, tracking, leading, or case for visual clarity, but avoid dense paragraphs that reduce legibility.\n*   **`role: \"script\"` Handling:** A `script` block, if present, is purely informational. It may inspire atmosphere, but must never be printed on the poster unless a `special_instructions` clause demands it.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** Any `special_instructions` field—either at `global_style` level or within individual `blocks`—acts as the **Creative Directive**.\n*   **Action:** Treat its content as an overriding brief when it conflicts with default behavior.\n    *   *Example:* If `global_style` suggests “Muted palette” but `special_instructions` calls for “Blazing red rays from the center,” prioritize the rays and adjust the palette accordingly.\n    *   Use this channel for surface enhancements such as screen‑print textures, grain, overprint effects, light flares, or other stylistic flourishes.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** `global_style` defines the canonical look that all posters in the same series must share.\n*   **Background:** Use the exact `background.color` and texture indicated. Every poster in the set should appear printed on the same stock.\n*   **Fonts:** Respect the specified typefaces: `font_palette.primary` for key slogans, `font_palette.secondary` for supporting copy.\n*   **Colors:** When a style references a `color_ref`, it must map precisely to the hex codes in `color_palette`.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`importance: \"dominant\"`) must share scale and anchoring across posters.\n    *   All supporting text or list blocks (`importance: \"supporting\"` or equivalent) must retain uniform typographic treatment.\n    *   All minor notes/captions (`importance: \"minor\"`) must use a consistent, smaller style.\n*   **Purpose:** This ensures that, when multiple posters are generated independently, they hang together as a cohesive campaign with matching hierarchy and rhythm.\n\n#### 5. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** Avoid repetitive “headline with centered bullets” arrangements. Each poster should feel like a unique print run in the same campaign.\n*   **Execution:**\n    *   **Copy‑heavy layouts:** Organize text into inventive structures (angled banners, split columns, offset blocks) rather than a simple vertical stack.\n    *   **Image‑driven layouts:** Let the imagery command attention; keep text minimal and strategically placed to support the visual story.\n    *   **Hybrid layouts:** Interlock text and imagery with overlaps, diagonals, and layered panels while preserving clarity.\n    *   **Whitespace:** Treat unprinted areas as deliberate design space. Do not stretch text or scale elements solely to “fill the page.”\n*   **Variety Techniques:**\n    *   Shift alignments (flush left, flush right, centered, or diagonal baselines).\n    *   Use varied banner shapes, badges, or ribbons as text containers.\n    *   Change visual weight from slide to slide (sometimes the title dominates, sometimes imagery or symbols take precedence).\n*   **Purpose:** The result should be a series of posters that feel energetic and collectible, unified by style but distinct in composition.\n\n### OUTPUT GENERATION\nRender a single propaganda-style poster image that obeys these directives.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Vintage Propaganda Poster\",\n    \"background\": {\n      \"color\": \"#F2E6D3\",\n      \"gradient\": null,\n      \"texture\": \"paper\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Impact\",\n      \"secondary\": \"Impact\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#8B4513\",\n      \"secondary\": \"#D2691E\",\n      \"text\": \"#1A1A1A\",\n      \"accent\": \"#C19A6B\"\n    },\n    \"special_instructions\": \"The poster emulates an aged propaganda broadside or carved woodcut print. It uses a bold black frame, an off-white parchment texture, and a restrained palette anchored in sepia, brick red, and midnight blue-grey. The composition is grid-based with ornamental Art Deco borders featuring stylized wheat and geometric motifs.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Dramatic, Historical, Artistic, Nostalgic\",\n    \"composition_style\": \"Grid-based, Symmetrical\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 73,\n      \"blocks\": [\n        {\n          \"id\": \"title-73\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"FROM BATTLEFIELD SIGNALS TO VOICES ON THE PAGE\",\n          \"style\": {\n            \"color\": \"#1A1A1A\",\n            \"font_family\": \"Impact\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center, spanning full width\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Headline set in uppercase, tightly tracked, centered inside a wide rectangular band with a thick black stroke and pale parchment fill.\"\n          }\n        },\n        {\n          \"id\": \"image-73-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_85_125_681_497.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_85_125_681_497.jpg\"\n              }\n            ],\n            \"source\": \"Violet Evergarden as a soldier in a battlefield setting\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-left quadrant\",\n            \"image_style\": \"woodcut print style, framed by black border\",\n            \"container\": \"rectangular frame\",\n            \"special_instructions\": \"Image depicts a blonde female soldier holding a rifle against a backdrop of mountains and a rising sun.\"\n          }\n        },\n        {\n          \"id\": \"image-73-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_697_126_1291_495.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_697_126_1291_495.jpg\"\n              }\n            ],\n            \"source\": \"Violet Evergarden as an Auto Memory Doll typing at a desk\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-right quadrant\",\n            \"image_style\": \"woodcut print style, framed by black border\",\n            \"container\": \"rectangular frame\",\n            \"special_instructions\": \"Image depicts the character typing on a typewriter with a dove carrying a letter in the background.\"\n          }\n        },\n        {\n          \"id\": \"body-73-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"FRONTLINES AND FIRE\",\n          \"style\": {\n            \"font_family\": \"Impact\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"dominant\",\n            \"position\": \"middle-left, below the soldier image\",\n            \"container\": \"panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Section label in uppercase, bold, centered within a beige rectangular placard bordered in black.\"\n          }\n        },\n        {\n          \"id\": \"body-73-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"- TRAINED AS A TOOL OF WAR\\n- CARRYING ORDERS THROUGH CHAOS\\n- EMERGING FROM CONFLICT SCARRED AND SEARCHING\",\n          \"style\": {\n            \"font_family\": \"Impact\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-left, below the 'FRONTLINES AND FIRE' header\",\n            \"container\": \"panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Bulleted phrases (dash-prefixed), uppercase, left-aligned inside the same beige panel system.\"\n          }\n        },\n        {\n          \"id\": \"body-73-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"LETTERS THAT CARRY HEARTS\",\n          \"style\": {\n            \"font_family\": \"Impact\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"dominant\",\n            \"position\": \"middle-right, below the writer image\",\n            \"container\": \"panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Section label in uppercase, bold, centered within a matching beige rectangle with a heavy black outline.\"\n          }\n        },\n        {\n          \"id\": \"body-73-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"- TRANSLATING SILENT FEELINGS INTO WORDS\\n- CONNECTING STRANGERS ACROSS DISTANCE\\n- LEARNING WHAT IT MEANS TO TREASURE SOMEONE\",\n          \"style\": {\n            \"font_family\": \"Impact\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"middle-right, below the 'LETTERS THAT CARRY HEARTS' header\",\n            \"container\": \"panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Dash-led bullet list, uppercase, neatly left-aligned within the beige content panel.\"\n          }\n        },\n        {\n          \"id\": \"footer-73\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"A PATH FROM ORDERS SHOUTED IN WAR TO WHISPERS PRESERVED IN INK.\",\n          \"style\": {\n            \"font_family\": \"Impact\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom, spanning full width\",\n            \"container\": \"banner\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Closing slogan in uppercase, centered within a low horizontal banner boxed by a thick black stroke along the poster’s base.\"\n          }\n        },\n        {\n          \"id\": \"illustration-73-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Vertical decorative border panels on the far left and right sides containing stylized wheat stalks and geometric Art Deco patterns in dark blue-grey and rust red.\",\n          \"style\": {\n            \"illustration_style\": \"decorative border pattern\",\n            \"color_scheme\": \"dark blue-grey, rust red, beige\",\n            \"position\": \"far left and far right edges, spanning vertically\",\n            \"special_instructions\": \"These ornaments enclose the central artwork and emphasize the retro propaganda character of the piece.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nGenerate the final propaganda poster in a widescreen 16:9 ratio, ensuring sharp lettering and a harmonious layout.", "input_images": ["images/multi_image/7e5bea6e-5963-42be-b2c3-21adbd375c16_input_0.jpg", "images/multi_image/7e5bea6e-5963-42be-b2c3-21adbd375c16_input_1.jpg"], "target_image": "images/multi_image/7e5bea6e-5963-42be-b2c3-21adbd375c16_target.png", "width": 1376, "height": 768}
+{"index": 222, "id": "39f90ad9-43f0-4150-828b-5eec79c84d2a", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, styled with a clean white background and thin charcoal borders, highlights a comparison of model performance. The title at the top reads \"TRUTHFULQA: MONOLINGUAL VS PLUG\" in bold, modern black letters. A short description underneath states: \"OVERVIEW: ACCURACY SCORES ACROSS LANGUAGES\".  \n\nThe slide is divided into two balanced sections.  \nOn the left, a compact text block titled \"LANGUAGES & SCORES:\" lists four bullet points:  \n• CHINESE  \n• KOREAN  \n• ITALIAN  \n• SPANISH  \n\nBelow the list, a brief note reads: \"PLUG SHOWS CONSISTENT GAINS OVER MONOLINGUAL MODELS.\"  \n\nOn the right, the sub image is prominently placed, showing grouped bars for each language. The light bars represent \"Monolingual\" scores, and the blue bars represent \"PLUG\" scores, with values ranging from about 44 to 61. The x‑axis is labeled \"TruthfulQA\" with the four languages, and the y‑axis is labeled \"TruthfulQA Score\".  \n\nA small caption beneath the sub image states: \"SUB IMAGE: VISUAL COMPARISON OF TRUTHFULQA SCORES.\"", "input_images": ["images/multi_image/39f90ad9-43f0-4150-828b-5eec79c84d2a_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 223, "id": "fe9ad55b-8ac8-4702-8e6e-62d33b355971", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 4. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Quiet Morning Read\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#F4F1EB\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #FFFFFF 0%, #F4F1EB 45%, #E3D9CF 100%)\\\",\\n      \\\"texture\\\": \\\"paper\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Playfair Display\\\",\\n      \\\"secondary\\\": \\\"Open Sans\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#C3341D\\\",\\n      \\\"secondary\\\": \\\"#3D4A4F\\\",\\n      \\\"text\\\": \\\"#2A2A2A\\\",\\n      \\\"accent\\\": \\\"#F0B35A\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"The slide should feel calm and reflective, like an early autumn morning in a courtyard café. Use a warm, off‑white paper texture background with very subtle grain. Accents pick up the deep red of the sweater and the muted greens from the foliage. Typography is modern serif for the title and clean sans-serif for supporting text. Keep plenty of negative space around the central image. Use soft drop shadows and slightly rounded corners to make the image panel feel inviting and tactile.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Calm, Reflective, Editorial\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 14,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-14\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Moments of Quiet Focus\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#3D4A4F\\\",\\n            \\\"font_family\\\": \\\"Playfair Display\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, above the image with generous margin\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"none\\\"\\n            },\\n            \\\"container\\\": null,\\n            \\\"special_instructions\\\": \\\"Title in mixed case, slightly letter-spaced, centered. Keep it compact so it does not compete with the image.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/fe9ad55b-8ac8-4702-8e6e-62d33b355971_input_0.png\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/fe9ad55b-8ac8-4702-8e6e-62d33b355971_input_0.png\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Person in a red sweater reading a newspaper at an outdoor table with coffee.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, occupying roughly 70% of slide width and 60% of height\\\",\\n            \\\"image_style\\\": \\\"full-color photo with soft drop shadow and subtle off-white frame\\\",\\n            \\\"container\\\": \\\"Thin border in #FFFFFF with outer line in #C3341D\\\",\\n            \\\"special_instructions\\\": \\\"Use slightly rounded corners (8–10px). Add a soft, diffuse shadow (#000000 at 10% opacity) to lift the image from the background. Keep the framing tight so the subject remains the clear focal point.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-14-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"A simple pause with coffee and headlines before the day begins.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Open Sans\\\",\\n            \\\"color\\\": \\\"#5A6468\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image frame\\\",\\n            \\\"container\\\": null,\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Set in smaller size, light to regular weight, centered; acts as a quiet subtitle without drawing attention away from the photograph.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-14-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two soft, overlapping circular gradient shapes echoing the red sweater and warm accent color, placed behind the lower left and upper right corners of the image.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"Uses #C3341D fading to transparent and #F0B35A fading to transparent\\\",\\n            \\\"position\\\": \\\"one circle at lower-left behind the image, another at upper-right behind the image\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (20–30%), large enough to peek out from behind the image edges and subtly frame it without overwhelming the photo.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nNow produce the final 16:9 slide image. Do not add extra text; prioritize clarity, hierarchy, and polished visual design.", "input_images": ["images/multi_image/fe9ad55b-8ac8-4702-8e6e-62d33b355971_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 224, "id": "cb2d8a1a-d3a8-43c9-a320-2eff3a5e4ff0", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are an AI slide-generation engine. Convert the supplied slide JSON into a high-quality presentation image.\n\nFollow these constraints exactly:\n- The JSON is authoritative. Render only the contents of `blocks`; never add new words.\n- Embed all source images into the slide.\n- `global_style` defines the shared visual identity. Preserve its fonts, colors, background, and style language.\n- `visual_hierarchy` describes what should feel dominant, secondary, or subtle.\n- `special_instructions` may refine or override the default look; respect it when present.\n\nDesign with variety. Do not default to a title plus bullet list. Use asymmetry, columns, modular cards, large focal visuals, restrained annotations, or generous negative space when appropriate. The final slide must feel clean, readable, and presentation-ready.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Scaling Performance Analytics\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#0F172A\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #1E293B 0%, #020617 70%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Roboto Condensed\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#38BDF8\\\",\\n      \\\"secondary\\\": \\\"#F97316\\\",\\n      \\\"text\\\": \\\"#E5E7EB\\\",\\n      \\\"accent\\\": \\\"#22C55E\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall feel is modern, analytical, and data-driven. Dark navy background with a subtle radial gradient to draw focus toward the center where the chart sits. Use crisp sans-serif typography and thin white gridlines if any additional elements are added. Minimal clutter, generous whitespace around the main figure, and subtle cyan accent lines to echo the plot colors. Lighting should feel like a softly lit dashboard screen with gentle glow around key elements, but remain professional and clean.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Analytical, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 12,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-12\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Performance vs. Training Tokens\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#E5E7EB\\\",\\n            \\\"font_family\\\": \\\"Roboto Condensed\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with generous margin above the image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Title in all-caps or small-caps, slightly expanded letter spacing for a dashboard-style look.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/cb2d8a1a-d3a8-43c9-a320-2eff3a5e4ff0_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/cb2d8a1a-d3a8-43c9-a320-2eff3a5e4ff0_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Line chart showing model performance across multiple benchmarks as training tokens increase.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning about 75% of slide width and 60% of height\\\",\\n            \\\"image_style\\\": \\\"clean embedded chart without additional annotations\\\",\\n            \\\"container\\\": \\\"subtle rounded rectangle with thin #1F2937 border\\\",\\n            \\\"special_instructions\\\": \\\"Apply a 10px border-radius to the container and a soft outer glow in #38BDF8 at low opacity to subtly highlight the chart as the focal element.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-12-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Long and short video QA steadily improve with scale, while hallucination drops as training tokens grow.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#CBD5F5\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, below the image with comfortable padding\\\",\\n            \\\"container\\\": \\\"rounded pill with semi-transparent #020617 background\\\",\\n            \\\"container_opacity\\\": \\\"0.7\\\",\\n            \\\"special_instructions\\\": \\\"Small, center-aligned sentence; use medium weight for readability without competing with the title.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-12-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A faint, oversized concentric ring pattern suggesting data waves, placed behind the chart area.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"centered behind the image block, extending slightly beyond its borders\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (around 0.15), using cyan (#38BDF8) and emerald (#22C55E) gradients to subtly echo the chart lines without distracting from the main figure.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nNow render the final slide as a polished 16:9 presentation image.", "input_images": ["images/multi_image/cb2d8a1a-d3a8-43c9-a320-2eff3a5e4ff0_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 225, "id": "07e6d1f5-5ebb-4371-8452-137be75242bd", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide, set against a soft cream-colored background, is titled \"CLIMATE MELODY MAKER\" in large, playful, multi-colored bubble letters at the top center; the word \"CLIMATE\" is styled in blue and green, while \"MELODY MAKER\" is in orange and red, flanked by small green leaf icons. The slide is organized into four vertical columns, each dedicated to a specific aspect of the topic, labeled \"RAIN,\" \"WIND,\" \"SUN,\" and \"SOUND\" respectively, using white text inside rounded banners matching their theme colors.\n\nThe first column, \"RAIN,\" features a blue banner with \"RAIN\" written in white uppercase letters. Below it, the section is titled \"SENSOR READINGS\" in bold black text. This section includes the sub-image of a blue frog character holding a tablet, with weather data displayed in a light blue box: \"21°C,\" \"85%,\" and \"Low,\" accompanied by icons for temperature, humidity, and rain. Below the image, the text \"SENSOR READINGS\" is repeated in bold, followed by the description: \"Gathering live atmospheric values for use as musical input.\"\n\nThe second column, \"WIND,\" is headed by a green banner with \"WIND\" in white. The sub-image here shows a yellow cat character next to a wind gauge and musical staves. The visual includes the text \"Wind\" and \"(decibels)\" alongside arrows indicating sound flow. Below this, the bold header \"PATTERN CONVERSION\" is followed by the text: \"Transforming motion of air into changes in loudness, register, and instrumental color.\"\n\nThe third column, \"SUN,\" is topped with an orange banner reading \"SUN\" in white. The sub-image depicts a sun shining over a musical staff. Below the image, the bold header \"MELODIC DESIGN\" is followed by three icons labeled \"Contour,\" \"Tempo,\" and \"Brightness.\" The description reads: \"Contour follows daily temperature shifts; Tempo responds to daylight duration; Brightness mirrors overall sky clarity.\"\n\nThe fourth column, \"SOUND,\" features a purple banner with \"SOUND\" in white. The sub-image shows a brown animal character wearing headphones in front of a house and speaker. The bold header \"IMMERSIVE PLAYBACK\" is followed by the text: \"Lets listeners explore changing climate conditions through evolving musical soundscapes.\"\n\nThe overall layout is clean and balanced, using a bright, cheerful color palette of blues, greens, oranges, and purples to distinguish the sections. Decorative elements like musical notes, sparkles, and clouds are scattered throughout to enhance the thematic connection between weather and music. Each column is enclosed in a rounded rectangular border corresponding to its color theme, creating a cohesive and visually engaging infographic.", "input_images": ["images/multi_image/07e6d1f5-5ebb-4371-8452-137be75242bd_input_0.jpg", "images/multi_image/07e6d1f5-5ebb-4371-8452-137be75242bd_input_1.jpg", "images/multi_image/07e6d1f5-5ebb-4371-8452-137be75242bd_input_2.jpg", "images/multi_image/07e6d1f5-5ebb-4371-8452-137be75242bd_input_3.jpg"], "target_image": "images/multi_image/07e6d1f5-5ebb-4371-8452-137be75242bd_target.png", "width": 1376, "height": 768}
+{"index": 226, "id": "d71d93d3-08d8-4114-b7b2-14ed52c255b4", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "Render one presentation slide from the provided Structured Content JSON.\n\nThe JSON is the single source of truth:\n1. Use `global_style` for the slide’s overall visual language.\n2. Use `visual_hierarchy` to guide emphasis, mood, and composition.\n3. Use `blocks` to determine all text and visual elements.\n4. Use `special_instructions` as the primary creative override when making design decisions.\n5. Use `source_images` only to embed images in the slide.\n\nYou may freely redesign the layout and typography, but do not alter, add, or remove any of the textual content contained in the JSON. Adjust only visual properties such as line breaks, alignment, weight, spacing, and scale to enhance readability.\n\nCreate a slide that feels custom-designed rather than based on a rigid template. Avoid repetitive bullet lists. Use negative space thoughtfully, group related elements clearly, align objects precisely, and ensure the composition reflects the informational structure.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Corporate Infographic\",\n    \"background\": {\n      \"color\": \"#F0F4FF\",\n      \"gradient\": \"radial-gradient(circle at center, #FFFFFF 0%, #E6EEFF 100%)\",\n      \"texture\": \"none\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#6C7BC9\",\n      \"secondary\": \"#4A90E2\",\n      \"text\": \"#2D3748\",\n      \"accent\": \"#FF6B6B\"\n    },\n    \"special_instructions\": \"The slide features a clean, modern corporate aesthetic with a soft blue gradient background. The design uses rounded rectangular containers with pastel color coding for different sections. Visual elements include 3D-style isometric illustrations and icons. The overall mood is professional yet approachable, with a focus on clarity and visual hierarchy.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Informative, Modern\",\n    \"composition_style\": \"Grid-based with centered title and five-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 603,\n      \"blocks\": [\n        {\n          \"id\": \"title-603\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"VALUING LOCAL NATIONALS\",\n          \"style\": {\n            \"color\": \"#6C7BC9\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"shadow\"\n              ],\n              \"outline_color\": null,\n              \"shadow_type\": \"drop\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, bold title text with a subtle drop shadow for depth\"\n          }\n        },\n        {\n          \"id\": \"illustration-603-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative blue circle with a gradient effect in the top-left corner\",\n          \"style\": {\n            \"illustration_style\": \"decorative background shape\",\n            \"color_scheme\": \"blue gradient\",\n            \"position\": \"top-left corner\",\n            \"special_instructions\": \"Partially visible, acts as a subtle background accent\"\n          }\n        },\n        {\n          \"id\": \"illustration-603-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A grid of small blue dots in the top-right corner\",\n          \"style\": {\n            \"illustration_style\": \"decorative pattern\",\n            \"color_scheme\": \"blue\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"3x3 grid of small circular dots\"\n          }\n        },\n        {\n          \"id\": \"illustration-603-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative blue arrow shape in the bottom-left corner\",\n          \"style\": {\n            \"illustration_style\": \"decorative shape\",\n            \"color_scheme\": \"blue gradient\",\n            \"position\": \"bottom-left corner\",\n            \"special_instructions\": \"Points towards the center of the slide\"\n          }\n        },\n        {\n          \"id\": \"illustration-603-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative orange cylinder with blue lines in the bottom-right corner\",\n          \"style\": {\n            \"illustration_style\": \"decorative shape\",\n            \"color_scheme\": \"orange and blue\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Abstract geometric decoration\"\n          }\n        },\n        {\n          \"id\": \"image-603-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_84_280_272_503.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_84_280_272_503.jpg\"\n              }\n            ],\n            \"source\": \"Cultural exchange illustration\",\n            \"caption\": \"CULTURAL INSIGHTS\"\n          },\n          \"style\": {\n            \"position\": \"first column, upper section\",\n            \"image_style\": \"contained within rounded blue panel\",\n            \"container\": \"rounded rectangular panel with blue border\",\n            \"special_instructions\": \"3D isometric illustration of two people exchanging a gift\"\n          }\n        },\n        {\n          \"id\": \"body-603-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Deep understanding of local customs, values, & traditions.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2D3748\",\n            \"importance\": \"supporting\",\n            \"position\": \"first column, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Left-aligned text, smaller font size\"\n          }\n        },\n        {\n          \"id\": \"image-603-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_316_275_548_557.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_316_275_548_557.jpg\"\n              }\n            ],\n            \"source\": \"Local knowledge illustration\",\n            \"caption\": \"LOCAL KNOWLEDGE\"\n          },\n          \"style\": {\n            \"position\": \"second column, upper section\",\n            \"image_style\": \"contained within rounded green panel\",\n            \"container\": \"rounded rectangular panel with green border\",\n            \"special_instructions\": \"3D isometric illustration with compass, lightbulb, and map\"\n          }\n        },\n        {\n          \"id\": \"body-603-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Expertise in geography, politics, economics, and infrastructure.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2D3748\",\n            \"importance\": \"supporting\",\n            \"position\": \"second column, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Left-aligned text, smaller font size\"\n          }\n        },\n        {\n          \"id\": \"image-603-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_828_275_1057_556.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_828_275_1057_556.jpg\"\n              }\n            ],\n            \"source\": \"Stakeholder relations illustration\",\n            \"caption\": \"STAKEHOLDER RELATIONS\"\n          },\n          \"style\": {\n            \"position\": \"fourth column, upper section\",\n            \"image_style\": \"contained within rounded blue panel\",\n            \"container\": \"rounded rectangular panel with blue border\",\n            \"special_instructions\": \"3D isometric illustration of people networking\"\n          }\n        },\n        {\n          \"id\": \"body-603-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Building trust with local communities, businesses, and authorities.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2D3748\",\n            \"importance\": \"supporting\",\n            \"position\": \"fourth column, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Left-aligned text, smaller font size\"\n          }\n        },\n        {\n          \"id\": \"image-603-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1082_278_1312_558.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1082_278_1312_558.jpg\"\n              }\n            ],\n            \"source\": \"Sustainability and growth illustration\",\n            \"caption\": \"SUSTAINABILITY & GROWTH\"\n          },\n          \"style\": {\n            \"position\": \"fifth column, upper section\",\n            \"image_style\": \"contained within rounded green panel\",\n            \"container\": \"rounded rectangular panel with green border\",\n            \"special_instructions\": \"3D isometric illustration of plant growth with coins\"\n          }\n        },\n        {\n          \"id\": \"body-603-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Fostering long-term development and positive impact.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2D3748\",\n            \"importance\": \"supporting\",\n            \"position\": \"fifth column, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Left-aligned text, smaller font size\"\n          }\n        },\n        {\n          \"id\": \"illustration-603-5\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A purple rounded panel containing speech bubbles with 'Hello!', 'Namaste', 'Hola' and a microphone icon\",\n          \"style\": {\n            \"illustration_style\": \"infographic element\",\n            \"color_scheme\": \"purple background with orange and blue speech bubbles\",\n            \"position\": \"third column, upper section\",\n            \"special_instructions\": \"Represents language and communication concept\"\n          }\n        },\n        {\n          \"id\": \"body-603-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Native language skills and nuance. Bridging cultural gaps.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2D3748\",\n            \"importance\": \"supporting\",\n            \"position\": \"third column, below illustration\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Left-aligned text, smaller font size\"\n          }\n        },\n        {\n          \"id\": \"caption-603-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"LOCAL NATIONALS CONTRIBUTIONS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#6C7BC9\",\n            \"position\": \"bottom-left\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"special_instructions\": \"Small text at the bottom of the slide\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nRender this slide in 16:9 format. Preserve the specified style, hierarchy, colors, typography, and exact visible text.", "input_images": ["images/multi_image/d71d93d3-08d8-4114-b7b2-14ed52c255b4_input_0.jpg", "images/multi_image/d71d93d3-08d8-4114-b7b2-14ed52c255b4_input_1.jpg", "images/multi_image/d71d93d3-08d8-4114-b7b2-14ed52c255b4_input_2.jpg", "images/multi_image/d71d93d3-08d8-4114-b7b2-14ed52c255b4_input_3.jpg"], "target_image": "images/multi_image/d71d93d3-08d8-4114-b7b2-14ed52c255b4_target.png", "width": 1376, "height": 768}
+{"index": 227, "id": "da1bcb0d-56cb-4590-af8c-627f41fa0d2c", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, styled with Peanuts characters, is titled \"MASTERING ACCRUAL BASICS\" in large blue text at the top. It features three panels arranged horizontally, corresponding to the provided sub-images. The first panel on the left is labeled \"1. KEY CONCEPT\" in a blue banner and shows Charlie Brown and Linus; the text below reads \"LOG REVENUE AS IT IS EARNED\" and \"LOG EXPENSES AS THEY ARISE,\" while a speech bubble from Linus states, \"Timing of cash doesn’t control the books!\" The middle panel is labeled \"2. RECORDING REVENUE\" and depicts Peppermint Patty and Snoopy with the text \"PROVIDE PRODUCTS/SERVICES,\" \"REVENUE IS CREATED,\" and \"ENTER SALES AT ONCE,\" alongside Snoopy's thought bubble saying \"Money in the making!\" The rightmost panel, labeled \"3. LINKING EXPENSES,\" shows Sally and Woodstock with the text \"SPEND TO GENERATE SALES\" and \"CAPTURE EXPENSES,\" plus a speech bubble stating, \"Tie each cost to the income it supports!\" and a box listing \"DIRECT COSTS (COGS)\" and \"OPERATING COSTS.\" At the bottom of the slide, a footnote reads \"*Peanuts characters and style pay homage to Charles M. Schulz*.\"", "input_images": ["images/multi_image/da1bcb0d-56cb-4590-af8c-627f41fa0d2c_input_0.jpg", "images/multi_image/da1bcb0d-56cb-4590-af8c-627f41fa0d2c_input_1.jpg", "images/multi_image/da1bcb0d-56cb-4590-af8c-627f41fa0d2c_input_2.jpg"], "target_image": "images/multi_image/da1bcb0d-56cb-4590-af8c-627f41fa0d2c_target.png", "width": 1376, "height": 768}
+{"index": 228, "id": "1e03e316-9963-4f35-ae8a-7f7fa387ebb8", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### MISSION\nTransform the input Structured Content JSON into one elegant, high-fidelity presentation slide.\n\n### STRICT CONTENT CONTROL\nOnly the JSON may supply visible content. Do not add explanatory copy, generated bullets, fake labels, placeholder text, or decorative words. If text is sparse, preserve the sparse feeling.\n\n### STYLE CONTROL\nHonor `global_style` as the governing design system. Use the specified background, color palette, typography, and overall visual mood. Apply `special_instructions` when provided, especially for atmosphere, lighting, texture, or layout emphasis.\n\n### BLOCK HANDLING\nRender visible blocks according to their role and hierarchy. Use `script` only as hidden context for the slide's theme and emotional tone. Keep text blocks readable and avoid cramped paragraphs.\n\n### IMAGE INTEGRITY & PRESERVATION\nWhen you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n\n### LAYOUT EXPECTATION\nDesign a slide with deliberate composition. Vary structure across slides: some may be visual-led, some text-led, some diagrammatic, some minimal. The output should feel premium, consistent, and ready for a professional deck.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Benchmark Comparison Highlight\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#0F172A\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #1E293B 0%, #020617 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"IBM Plex Sans\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#38BDF8\\\",\\n      \\\"secondary\\\": \\\"#A5B4FC\\\",\\n      \\\"text\\\": \\\"#E5E7EB\\\",\\n      \\\"accent\\\": \\\"#FACC15\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall look is modern and data-centric, with a dark dashboard-style background that makes the white table image pop. Subtle glow behind the image, thin neon-like accent lines, and clean sans-serif typography emphasize clarity and precision. No heavy textures; everything feels crisp, minimal, and analytical.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Minimalist, Data-Driven\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 9,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-9\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Video QA Benchmarks at a Glance\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#E5E7EB\\\",\\n            \\\"font_family\\\": \\\"IBM Plex Sans\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with generous margin above the image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Title is in sentence case, wide letter-spacing, centered horizontally.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/1e03e316-9963-4f35-ae8a-7f7fa387ebb8_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/1e03e316-9963-4f35-ae8a-7f7fa387ebb8_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Table comparing three video question-answering benchmarks with durations, number of videos, and QA type.\\\",\\n            \\\"caption\\\": \\\"Key dataset statistics summarizing MVBench, Egoschema, and VideoMME.\\\"\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center of slide, spanning about 70% of the width and slightly above vertical midpoint\\\",\\n            \\\"image_style\\\": \\\"soft outer glow using primary color and a thin light border\\\",\\n            \\\"container\\\": \\\"subtle rounded rectangle with translucent dark overlay\\\",\\n            \\\"special_instructions\\\": \\\"Apply a 12px border-radius, 1px border in #38BDF8 at 40% opacity, and a soft drop shadow to lift the table from the dark background.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-9-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Short clips to multi-minute videos: coverage from 16 seconds to 17 minutes, all with multi-choice QA.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#A5B4FC\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image\\\",\\n            \\\"container\\\": null,\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Small, centered text; keep it on a single line where possible.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-9-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two thin, semi-transparent cyan and indigo concentric rounded rectangles behind the image, slightly larger than the image frame, suggesting a glowing data dashboard.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"centered behind the image block\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (around 25%), soft edges, no hard borders, layered so they do not compete with the table but add depth.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate the slide image as a clean widescreen presentation frame. Respect `global_style`, `visual_hierarchy`, `blocks`, and any `special_instructions`.", "input_images": ["images/multi_image/1e03e316-9963-4f35-ae8a-7f7fa387ebb8_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 229, "id": "73b372ed-0dd5-4a55-bd45-2a4ee9db7ad0", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are given a JSON description of one slide. Produce the final slide image with professional presentation quality.\n\nInterpretation rules:\n- `global_style` sets the exact design constraints and must be followed closely.\n- `blocks` contain all visible content. Do not create extra text.\n- `visual_hierarchy` tells you how to prioritize scale, position, contrast, and attention.\n- `special_instructions` should be treated as explicit art direction.\n- `source_images` are the images to embed in the slide.\n\nRendering rules:\n- Keep typography sharp, readable, and inside the canvas.\n- Use the provided colors and fonts consistently.\n- Prefer intentional whitespace over filler.\n- Adapt the layout to the material instead of using the same template every time.\n- Ensure the slide looks like part of a cohesive deck while still having its own visual identity.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Cute Educational Infographic\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#D3E8F5\",\n      \"secondary\": \"#1A1A1A\",\n      \"text\": \"#1A1A1A\",\n      \"accent\": \"#FFD700\"\n    },\n    \"special_instructions\": \"The slide features a playful, cartoon-style aesthetic with cute animal characters (bears, cats, rabbits) and medical professionals. The design uses soft pastel colors, rounded shapes, and decorative elements like stars and hearts. The overall mood is friendly and approachable, making complex medical information accessible. Visual elements are arranged in a clear left-to-right flow with arrows indicating progression.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Playful, Educational, Friendly\",\n    \"composition_style\": \"Three-column grid with horizontal flow\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 266,\n      \"blocks\": [\n        {\n          \"id\": \"title-266\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"HOW DOES A VACCINE GET RECOMMENDED?\",\n          \"style\": {\n            \"color\": \"#1A1A1A\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Title is contained within a light blue rounded rectangular banner with a magnifying glass icon on the left and decorative stars/hearts on the right\"\n          }\n        },\n        {\n          \"id\": \"illustration-266-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Magnifying glass icon in dark grey/black positioned to the left of the title banner\",\n          \"style\": {\n            \"illustration_style\": \"flat icon\",\n            \"color_scheme\": \"dark grey/black\",\n            \"position\": \"top-left of title banner\",\n            \"special_instructions\": \"Simple line art style magnifying glass\"\n          }\n        },\n        {\n          \"id\": \"illustration-266-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative yellow stars and pink hearts scattered around the title banner\",\n          \"style\": {\n            \"illustration_style\": \"decorative elements\",\n            \"color_scheme\": \"yellow and pink\",\n            \"position\": \"around title banner\",\n            \"special_instructions\": \"Small decorative elements adding playful touch\"\n          }\n        },\n        {\n          \"id\": \"body-266-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"STEP 1: STUDY THE VACCINE\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"dominant\",\n            \"position\": \"left column, above first image\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold uppercase heading for first column\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_69_235_470_571.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_69_235_470_571.jpg\"\n              }\n            ],\n            \"source\": \"Gathering Information illustration\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left column, center\",\n            \"image_style\": \"cartoon illustration with cute animals\",\n            \"container\": null,\n            \"special_instructions\": \"Shows cute animals collecting data with telescope, tablet, and documents\"\n          }\n        },\n        {\n          \"id\": \"body-266-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"STEP 1: STUDY THE VACCINE\\nReviewing lab and clinical trial data to see how well the vaccine works. Looking closely at safety signals and side effects.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"First line is bold uppercase heading, followed by two lines of descriptive text\"\n          }\n        },\n        {\n          \"id\": \"body-266-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"STEP 2: WEIGH RISKS AND BENEFITS\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"dominant\",\n            \"position\": \"center column, above second image\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold uppercase heading for second column\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_493_249_904_565.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_493_249_904_565.jpg\"\n              }\n            ],\n            \"source\": \"Evaluation and Analysis illustration\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center column, center\",\n            \"image_style\": \"cartoon illustration with cute animals\",\n            \"container\": null,\n            \"special_instructions\": \"Shows cute animals at control panel with graphs and scales weighing cost vs benefit\"\n          }\n        },\n        {\n          \"id\": \"body-266-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"STEP 2: WEIGH RISKS AND BENEFITS\\nComparing protection from disease with possible harms. Considering costs, supplies, and impact on different groups.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"center column, bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"First line is bold uppercase heading, followed by two lines of descriptive text\"\n          }\n        },\n        {\n          \"id\": \"body-266-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"STEP 3: MAKE PUBLIC HEALTH GUIDANCE\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"dominant\",\n            \"position\": \"right column, above third image\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold uppercase heading for third column\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_931_243_1337_566.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_931_243_1337_566.jpg\"\n              }\n            ],\n            \"source\": \"Decision Making and Recommendations illustration\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right column, center\",\n            \"image_style\": \"cartoon illustration with cute animals and medical professionals\",\n            \"container\": null,\n            \"special_instructions\": \"Shows cute animals with thumbs up, approval stamp, and happy medical staff with public health sun\"\n          }\n        },\n        {\n          \"id\": \"body-266-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"STEP 3: MAKE PUBLIC HEALTH GUIDANCE\\nHealth experts issue recommendations on who should get the vaccine and when. Governments use this advice to plan programs.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, bottom\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"First line is bold uppercase heading, followed by two lines of descriptive text\"\n          }\n        },\n        {\n          \"id\": \"illustration-266-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Curved blue arrows connecting the three columns from left to right, indicating process flow\",\n          \"style\": {\n            \"illustration_style\": \"flowchart arrows\",\n            \"color_scheme\": \"light blue\",\n            \"position\": \"between columns, upper section\",\n            \"special_instructions\": \"Two curved arrows showing progression from Gathering to Evaluation to Decision Making\"\n          }\n        },\n        {\n          \"id\": \"illustration-266-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Straight blue arrows connecting the three text sections at the bottom from left to right\",\n          \"style\": {\n            \"illustration_style\": \"flowchart arrows\",\n            \"color_scheme\": \"light blue\",\n            \"position\": \"between bottom text blocks\",\n            \"special_instructions\": \"Two straight arrows showing progression between the three descriptive text sections\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nCreate the final 16:9 slide image from the JSON above, following all style and content rules exactly.", "input_images": ["images/multi_image/73b372ed-0dd5-4a55-bd45-2a4ee9db7ad0_input_0.jpg", "images/multi_image/73b372ed-0dd5-4a55-bd45-2a4ee9db7ad0_input_1.jpg", "images/multi_image/73b372ed-0dd5-4a55-bd45-2a4ee9db7ad0_input_2.jpg"], "target_image": "images/multi_image/73b372ed-0dd5-4a55-bd45-2a4ee9db7ad0_target.png", "width": 1376, "height": 768}
+{"index": 230, "id": "a357a6e6-c5f5-4b66-8397-02d3481bd75d", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The slide presents a stylized, vintage-inspired infographic titled **\"ENCRYPTION: GUARDING YOUR DIGITAL FRONTIER,\"** rendered in a retro poster aesthetic with muted greens, warm yellows, deep reds, and cream tones. The entire composition is enclosed within a thin, double-lined border in dark olive green, giving it the feel of an old adventure map or propaganda poster. At the top, centered horizontally, the title appears in large, bold, all-caps serif font — **“ENCRYPTION: GUARDING YOUR DIGITAL FRONTIER”** — set against a cream-colored banner that spans the width of the slide, just beneath the upper border.  \n\nBelow this header lies the central visual narrative: a lone explorer in khaki safari attire, complete with pith helmet and boots, strides confidently from left to right across a rolling landscape. He carries a shield emblazoned with a keyhole emblem, symbolizing protection. To his left, shadowy, claw-like hands reach out from off-frame, suggesting lurking threats. Beneath his feet, a winding path marked with dotted lines and colored dots labeled “data” flows like a river toward the right side of the scene. Along this path, encrypted text strings such as “ARP1B5L-D75U-VIR8QK CH4L35T8Q00P” and “G71L-P7AB.J11L0RB.JQRB.5018877” appear, visually representing scrambled information. In the background, a red-and-white striped lighthouse stands atop a hill, its beam cutting through the sky; atop the lighthouse is a small sign reading “KEY” above a golden key icon. To the right, a stone archway inscribed with “SAFE ZONE” frames a large golden padlock hanging at its center, guarding the entrance to security. Beyond the arch, swirling ribbons of cryptic symbols in teal and maroon suggest encoded data streams flowing into safety.  \n\nBeneath this main illustration, three distinct icon-text groupings are arranged horizontally. On the left, a shield icon with a keyhole contains the heading **“LOCK & DEFEND”** in bold, uppercase serif font, followed by the descriptive line **“Transforms plain details into sealed cipher text”** in smaller, regular-weight sans-serif type. In the center, interlocking gears with a padlock motif accompany the heading **“PROTECT CONFIDENTIALITY”** in matching bold serif, with the subtext **“Keeps personal and business information out of hostile hands”** below. On the right, two stylized human figures shaking hands over a treasure chest topped with a padlock introduce the heading **“SECURE CONNECTIONS”** in the same bold serif style, with the explanation **“Supports verified identities, safe sharing, and reliable transactions”** underneath.  \n\nAt the very bottom, centered beneath all three sections, the tagline **“EMBRACE ENCRYPTION. TRAVEL THE NET WITH CONFIDENCE.”** appears in all-caps serif font, slightly smaller than the main headings but larger than the body text, anchoring the slide’s motivational message. The sub-images provided correspond to cropped segments of this final slide: one captures the left portion including the explorer and data path, another isolates the lighthouse and safe zone arch, and the third focuses on the three bottom icons and their associated text blocks — each serving as visual input for reconstructing the full slide’s layout and content. No charts, tables, or flowcharts are present; the entire design relies on symbolic imagery and typographic hierarchy to convey its message.", "input_images": ["images/multi_image/a357a6e6-c5f5-4b66-8397-02d3481bd75d_input_0.jpg"], "target_image": "images/multi_image/a357a6e6-c5f5-4b66-8397-02d3481bd75d_target.png", "width": 1376, "height": 768}
+{"index": 231, "id": "b86d2753-a2e0-437e-a822-aa80b9a74717", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Featuring a clean, structured layout against a warm cream-colored background, this presentation slide is titled **“SUNLIT SERENITY”** in large, bold, maroon serif font centered at the top, underscored by a thin horizontal maroon line stretching nearly the full width of the slide.  \n\nDirectly beneath the title, a single, dominant visual panel is centered on the slide, making the imagery the primary focus. This wide rectangular frame holds **the sub image**: a woman reclining on a blanket in a golden field at sunset, wearing a delicate, lace-patterned light bodysuit. She is captured in a relaxed pose, propped on one arm, with the other arm extended upward, holding a cluster of green grapes toward her lips. Soft sunlight from the right side of the frame creates a warm halo along her arm and hair, while long grasses in shades of amber and olive surround her. In the foreground, on the white blanket, a shallow plate displays an arrangement of fruit — green grapes, bright apples, and a pear — adding a subtle pop of color that harmonizes with the sunset sky of muted gold and blue. The composition feels intimate yet airy, with generous negative space in the sky above the figure.\n\nA thin maroon border encloses the image, visually separating it from the background and reinforcing the structured aesthetic. Subtle drop shadowing behind the panel gives a sense of depth without overwhelming the minimal design.\n\nCentered just below the image, a short heading in bold, dark brown sans-serif font reads **“Golden Hour Escape”**. Directly under this heading, a single concise line of text in regular-weight, dark brown sans-serif font states:  \n**“An intimate moment of calm, framed by light, nature, and simple pleasures.”**\n\nAll text and visual elements are aligned along a central vertical axis, with equal padding above and below the image panel, preserving a balanced, uncluttered look. The restrained color palette of cream, maroon, dark brown, and the natural hues present within the sub image emphasizes tranquility and elegance, allowing the photograph to remain the centerpiece of the slide.", "input_images": ["images/multi_image/b86d2753-a2e0-437e-a822-aa80b9a74717_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 232, "id": "cfa3eb4b-de65-46da-8746-974ad128d74c", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide is titled **\"VIDEO UNDERSTANDING BENCHMARKS\"** in large gold serif text centered at the top. The layout is split into two vertical panels against a dark background, separated by a thin vertical line. The left panel features the sub image displaying a performance comparison table; above this image is the heading **\"BASELINE MODEL\"**, and below it are the lines **\"QWEN2-VL-7B METRICS\"** and **\"CORE CAPABILITIES SNAPSHOT\"**. The right panel displays the same sub image of the benchmark table; this section is headed **\"FINE-TUNED MODELS\"**, with the text **\"ORIGINAL VS RECAPTION FT\"** and **\"IMPROVED SCORES ACROSS TASKS\"** positioned underneath. All text elements share the same gold serif styling.", "input_images": ["images/multi_image/cfa3eb4b-de65-46da-8746-974ad128d74c_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 233, "id": "39fa250c-488c-4243-8bd7-4d90860eaf3e", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Headlined by a bold declaration of the raw strength and quiet grace of wildlife in winter, this presentation slide is set on a clean, soft off‑white background, divided vertically into two balanced sections by a thin, cool gray line. The left side is dedicated to minimal, impactful text, while the right side is reserved for the visual presence of the sub image.\n\nOn the top left, the title **\"Solitude of a Tiger in Snow\"** appears in a large, deep forest‑green serif font. Directly beneath, a concise sub‑headline in the same color but slightly smaller bold serif typeface reads: **\"A lone predator moves silently through a frozen world.\"** \n\nA bit lower in the left column, two brief lines of regular‑weight forest‑green text add just enough context while keeping attention on the sub image:\n- **\"Every step leaves a fleeting mark on untouched snow.\"**  \n- **\"Power, focus, and resilience define survival in the cold.\"**\n\nOn the right side of the slide, the sub image is prominently embedded, occupying nearly the full height of the pane with narrow margins to emphasize immersion. The tiger is centered horizontally, striding forward through a snowy landscape that fills the background with soft whites and pale grays. Fine snowflakes are visible against the animal’s orange coat, its dark stripes and intense gaze drawing the viewer’s eye immediately. The image alignment ensures the tiger’s face is roughly at the viewer’s eye level, reinforcing a sense of direct encounter.\n\nBelow the image, in a small, light gray sans‑serif caption aligned to the center of the right column, a single understated line reads: **\"Sub image: A tiger advancing through deep winter snow.\"**\n\nThe overall slide composition mirrors the example structure: calm, spacious margins, restrained color palette, and a strong visual hierarchy that makes the sub image the dominant focal point while the minimal text quietly frames its mood and meaning.", "input_images": ["images/multi_image/39fa250c-488c-4243-8bd7-4d90860eaf3e_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 234, "id": "d3cefdbe-cc1f-4fbf-b0c7-1251c8685b19", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The slide presents a concise visual explanation of modern sequence modeling architectures, framed within a clean, technology‑focused layout on a soft off‑white background that covers the entire canvas. A thin, dark gray border with slightly rounded corners encloses the composition, giving the slide a polished, poster‑like appearance.\n\nAt the very top center, a wide, deep‑blue rounded rectangular banner stretches almost from edge to edge. Inside this banner, a short, minimal title is set in bold, white sans‑serif text:  \n**“Comparing Sequence Mixers”**  \nThe lettering is centered and evenly spaced, keeping the title compact but visually prominent. Just above the right edge of the banner, three short radiating strokes in dark gray arc outward, suggesting emphasis and drawing attention toward the headline.\n\nDirectly below the banner, the slide is gently divided into two conceptual zones: a compact text column on the left and a broad image display on the right, though no hard divider line is drawn. The intent is for the sub image to dominate, with text acting merely as light annotation.\n\nOn the left side, aligned vertically and centered with respect to the slide height, are two small white rounded rectangles that serve as caption boxes. Each box is outlined in a light gray stroke and filled with subtle shadow to lift it slightly off the background.\n\n- The top box contains a brief key phrase in dark blue bold sans‑serif type:  \n  **“Softmax Attention: token‑wise interaction via Q, K, V.”**  \n  The text is centered in a single, compact line.\n\n- The lower box, positioned with a narrow gap beneath the first, reads in similar style:  \n  **“Hedgehog & Mamba: efficient mixers for long sequences.”**\n\nThese two boxes provide just enough context to orient the viewer without competing with the main visual content.\n\nOccupying the entire right half of the slide—and extending slightly into the central vertical region—is the sub image itself. It is inset within a thin, charcoal gray frame that matches the border of the slide, preserving its original rectangular proportions. The sub image is placed so that its center aligns horizontally with the slide’s midline, ensuring it feels like the true focal point.\n\nWithin this framed sub image, several architectural diagrams are visible:\n\n- In the upper‑right quadrant of the sub image, a block diagram highlights a **Softmax Attention** stack. A green‑tinted rounded rectangle subtly emphasizes the classic attention mechanism: linear projections producing Q and K, their matrix product feeding a softmax operation, and the resulting weights multiplying V. Norm, MLP, and residual connectors encircle this region.\n\n- In the lower‑left quadrant, a **Hedgehog Attention** module appears inside a blue‑tinted box. Here, φ‑transformed queries and keys are multiplied via φ(Q)φ(K)ᵀ, indicating a kernelized or linear attention variant. The broader encoder stack remains visible, with Norm and MLP layers framing the attention block.\n\n- In the central‑lower region, a **HedgeMamba** architecture merges elements of Hedgehog attention and a state space model (SSM). A blue‑highlighted portion again marks φ(Q)φ(K)ᵀ, while a yellow‑tinted region emphasizes a parallel SSM pathway, including a convolutional block and a gate, suggesting hybrid mixing of token interactions.\n\n- On the far right of the sub image, a **Mamba Mixer** schematic is set within a yellow‑shaded box. It shows a linear projection producing parameters B and C, a diagonal Λ, and their combination Cᵀ B ⊙ Λ, with an accompanying convolution and gate. Residual arrows and normalization blocks complete the pipeline.\n\nThe slide’s background around the sub image is lightly decorated with understated, technology‑themed doodles drawn in fine dark‑gray lines: a small circuit‑like branching node in the upper left corner, a pair of minimal wave‑forms near the lower right, and a dashed rectangular outline suggesting a “module” near the bottom center. These accents remain subtle and are kept in monochrome so they do not distract from the sub image.\n\nAt the very bottom, centered under both text and image, a thin line of small gray sans‑serif text serves as an unobtrusive footer:  \n**“Visual summary of attention, Hedgehog, and Mamba‑style sequence layers.”**  \n\nOverall, the slide maintains minimal textual content while clearly spotlighting the sub image, using color accents and restrained annotations to emphasize the comparison of different sequence mixers without crowding the visual space.", "input_images": ["images/multi_image/d3cefdbe-cc1f-4fbf-b0c7-1251c8685b19_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 235, "id": "bc140b61-2188-4ee0-9fb6-eaa39dd0535f", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "Render one presentation slide from the provided Structured Content JSON.\n\nThe JSON is the single source of truth:\n1. Use `global_style` for the slide's visual system.\n2. Use `visual_hierarchy` for emphasis, mood, and composition.\n3. Use `blocks` for exact text and visual elements.\n4. Use `special_instructions` as a priority creative override.\n5. Use `source_images` to embed in the slide.\n\nNever add or rewrite textual content. You may only improve typography through layout decisions such as line breaks, alignment, font weight, spacing, and scale.\n\nMake the slide feel designed, not templated. Avoid monotonous bullet layouts. Use thoughtful whitespace, clean grouping, precise alignment, and a composition that matches the content.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Neural Sequence Learning Blueprint\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#0B1220\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #1F2937 0%, #020617 65%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": \\\"subtle faint grid of thin lines in #111827\\\"\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Montserrat\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#38BDF8\\\",\\n      \\\"secondary\\\": \\\"#F97316\\\",\\n      \\\"text\\\": \\\"#E5E7EB\\\",\\n      \\\"accent\\\": \\\"#22C55E\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall feel is a modern, technical research slide suited for a machine learning presentation. Dark, slightly bluish background with a soft radial glow gives a focused, lab-like atmosphere. Fine grid lines in the backdrop suggest structure and data. Typography is clean, geometric sans-serif with generous tracking. Accents in cyan, warm orange, and green highlight key elements without overpowering the central figure. Minimal additional decoration to keep attention on the image.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Technical, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 26,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-26\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Diverse Distillation for Sequence Models\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#E5E7EB\\\",\\n            \\\"font_family\\\": \\\"Montserrat\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with comfortable margin from top edge\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"none\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Title is set in large, all-caps letters with slightly increased letter spacing. A thin cyan underline (#38BDF8) spans about 60% of the title width, positioned just below the text to subtly connect it visually to the diagram below.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/bc140b61-2188-4ee0-9fb6-eaa39dd0535f_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/bc140b61-2188-4ee0-9fb6-eaa39dd0535f_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Diagram illustrating diverse distillation with reference selection and model outputs.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center of slide, occupying roughly 70% of width and 55% of height\\\",\\n            \\\"image_style\\\": \\\"clean, no inner effects, subtle outer glow\\\",\\n            \\\"container\\\": \\\"rounded rectangle frame with thin #38BDF8 border\\\",\\n            \\\"special_instructions\\\": \\\"Place the image prominently in the center. Use a 12px border radius on the container and a thin 1.5px cyan border. Add a very soft outer shadow (spread 8px, blur 24px, color rgba(0,0,0,0.6)) to lift it off the dark background.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-26-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Student distributions are distilled from multiple reference sequences, balancing accuracy with diversity.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#9CA3AF\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image frame\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Single-line sentence in small, medium-weight text, centered. Keep ample negative space around it to maintain the image as focal point.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"caption-26-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"caption\\\",\\n          \\\"content\\\": \\\"Loss is computed against the best reference among diverse candidate sequences.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#F97316\\\",\\n            \\\"position\\\": \\\"bottom-left, aligned with left edge of image container\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"normal\\\",\\n              \\\"text_effect\\\": [\\n                \\\"none\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"none\\\"\\n            },\\n            \\\"special_instructions\\\": \\\"Caption set in small caps with slightly condensed tracking. Positioned close to the lower-left corner of the image to feel like a technical note.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-26-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two semi-transparent concentric circles in cyan and green behind the central image, hinting at probability distributions.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"centered behind the image block, extending slightly beyond its edges\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (around 0.20), smooth gradients from #38BDF8 to #22C55E, blurred edges. Shapes should stay subtle and never distract from legibility of the image.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate the slide image as a clean widescreen presentation frame. Respect `global_style`, `visual_hierarchy`, `blocks`, and any `special_instructions`.", "input_images": ["images/multi_image/bc140b61-2188-4ee0-9fb6-eaa39dd0535f_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 236, "id": "fcc0a77b-70e3-41ac-8037-d28dcc1aa3f0", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Against a soft cream background with faint, curved line patterns suggesting the surface of a table, this presentation slide is titled **\"CELEBRATING SIMPLE SANDWICHES\"** in large, bold, dark brown sans‑serif capital letters at the upper left. The layout is split visually into a left area dominated by the sub image and a narrower right area with minimal supporting text and icons.\n\nOn the left half, the **sub image** fills most of the space: a close-up photograph of a plate holding a thick, golden-brown sandwich cut into triangles. The bread looks soft and slightly toasted, with a creamy filling just visible between the slices. Beside the sandwich, fresh green leaves add color and texture. The background is blurred, keeping the viewer’s focus on the warm tones of the bread and the crisp freshness of the greens. A thin, rounded rectangular frame in light brown subtly encloses the photo, making it feel like a featured image on a menu.\n\nTo the right of the **sub image**, the text begins with the heading **\"FOCUS OF THIS SLIDE\"** in bold, dark brown sans‑serif font. Below, in smaller regular black sans‑serif text, a single descriptive line reads:  \n- *A comforting sandwich, served simply, highlighting texture, warmth, and freshness.*\n\nA narrow horizontal rule in a muted beige separates this short description from a lower text block titled **\"VISUAL NOTES\"** in bold, dark brown font. Under this are three concise bullet points in regular black text, each preceded by a small circular icon:\n\n- A tiny bread icon in soft tan, followed by: *Soft, thick slices with lightly toasted edges.*  \n- A small plate icon outlined in gray, followed by: *White plate framing the sandwich as the centerpiece.*  \n- A delicate leaf icon in green, followed by: *Fresh greens for color contrast and balance.*\n\nAt the bottom right corner of the slide, a small caption in light gray italics reads: *Sub image: close-up view of a simple sandwich on a plate, inviting attention to its texture and warmth.*\n\nOverall, the design is calm and uncluttered, ensuring that the **sub image** of the sandwich remains the main visual focus, with the surrounding text and icons offering just enough context to guide the viewer’s appreciation.", "input_images": ["images/multi_image/fcc0a77b-70e3-41ac-8037-d28dcc1aa3f0_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 237, "id": "9212a726-a774-4503-8c32-6016e99dd5b4", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "The slide has a cream-colored background and is organized around a prominent text block paired with a line chart. At the top, the heading reads \"Sustainable Urban Development\" in a large, reddish-brown serif font. Beneath the title is a paragraph of body text in the same color: \"Modern cities face the challenge of balancing population growth with environmental resilience and social well-being. Sustainable urban development integrates green infrastructure, efficient public transportation, and energy‑conserving building design to minimize ecological footprints. By leveraging data-driven planning, community participation, and innovative technologies—such as smart grids, mixed‑use zoning, and low‑carbon transit systems—planners can create neighborhoods that are livable, inclusive, and adaptable. This approach supports long-term economic vitality, reduces pollution and resource consumption, and enhances quality of life for diverse urban populations.\" The bottom half displays the line graph from the sub-images, plotting data across a vertical axis marked 0 to 40 and a horizontal axis labeled \"Item 1\" through \"Item 5.\" The chart features two lines in reddish-brown hues: a longer, darker line tracking a general upward trend peaking near 40, and a shorter, lighter line rising before dipping and recovering.", "input_images": ["images/multi_image/9212a726-a774-4503-8c32-6016e99dd5b4_input_0.jpg"], "target_image": "images/multi_image/9212a726-a774-4503-8c32-6016e99dd5b4_target.png", "width": 1376, "height": 768}
+{"index": 238, "id": "7d9fe018-2044-45e9-baa2-9914cb493888", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are given a JSON description of one slide. Produce the final slide image with professional presentation quality.\n\nInterpretation rules:\n- `global_style` sets the exact design constraints and must be followed closely.\n- `blocks` contain all visible content. Do not create extra text.\n- `visual_hierarchy` tells you how to prioritize scale, position, contrast, and attention.\n- `special_instructions` should be treated as explicit art direction.\n- `source_images` are the images to embed in the slide.\n\nRendering rules:\n- Keep typography sharp, readable, and inside the canvas.\n- Use the provided colors and fonts consistently.\n- Prefer intentional whitespace over filler.\n- Adapt the layout to the material instead of using the same template every time.\n- Ensure the slide looks like part of a cohesive deck while still having its own visual identity.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Playful Anonymity\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#FFE5EC\\\",\\n      \\\"gradient\\\": \\\"radial gradient from #FFE5EC in the center to #FFD0DC at the edges\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Futura\\\",\\n      \\\"secondary\\\": \\\"Helvetica Neue\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#E53952\\\",\\n      \\\"secondary\\\": \\\"#4CA2B8\\\",\\n      \\\"text\\\": \\\"#2B2B2B\\\",\\n      \\\"accent\\\": \\\"#FFFFFF\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"The slide should feel modern, minimal, and slightly surreal to match the whimsical character with the red box on their head. Use clean vector-style elements with soft, even lighting. Plenty of negative space surrounds the central image. Curved shapes and subtle gradients keep the look playful but refined, emphasizing contrast between the cool blue coat and warm red accessories.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Playful, Minimalist, Surreal\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 22,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-22\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"The Art of Standing Out Anonymously\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#E53952\\\",\\n            \\\"font_family\\\": \\\"Futura\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, above the main image with generous spacing\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"none\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Title in all caps or small caps, tracking slightly expanded, centered. Keep it clean with no box behind it so the background gradient shows through.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/7d9fe018-2044-45e9-baa2-9914cb493888_input_0.png\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/7d9fe018-2044-45e9-baa2-9914cb493888_input_0.png\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"A person in a teal coat with red gloves and a red box over their head, holding a jar with a straw against a pink background.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center of the slide, occupying roughly 65–70% of the slide height\\\",\\n            \\\"image_style\\\": \\\"no border, soft drop shadow\\\",\\n            \\\"container\\\": \\\"subtle rounded-rectangle frame with transparent fill and thin #4CA2B8 stroke\\\",\\n            \\\"special_instructions\\\": \\\"Place the image on a slightly oversized rounded rectangle (corner radius about 32px) that lightly frames it without cropping. Add a very soft, wide drop shadow (#E08A9A at 25% opacity) behind the image container to lift it from the background.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-22-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Bold colors, hidden face, unforgettable presence.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Helvetica Neue\\\",\\n            \\\"color\\\": \\\"#2B2B2B\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image frame\\\",\\n            \\\"container\\\": null,\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Short sentence centered, medium weight, slightly smaller than the title. Maintain ample margin from slide bottom for a clean, airy feel.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-22-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two large, soft gradient circles partially peeking from behind the image: one teal circle on the lower left and one red-pink circle on the upper right.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"layered behind the central image container, extending beyond its edges\\\",\\n            \\\"special_instructions\\\": \\\"Circles should be low-opacity (around 25–30%), with smooth gradients from #4CA2B8 to #8DD2E3 on the left and from #E53952 to #FF8A9A on the right. They should not distract from the image but add depth and echo the colors of the coat and gloves.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate the slide image as a clean widescreen presentation frame. Respect `global_style`, `visual_hierarchy`, `blocks`, and any `special_instructions`.", "input_images": ["images/multi_image/7d9fe018-2044-45e9-baa2-9914cb493888_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 239, "id": "747c5d7b-50a3-4101-a61e-27bbcb2485ba", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide, titled \"How Do Indian Philosophies Map the Journey of Life?\" in a large, purple, serif font at the top center, uses a soft, pastel gradient background that shifts from light blue to lavender, accented by abstract, iridescent purple and pink ribbons in the corners. The slide’s primary content is arranged in a split layout: a diagram on the left and three explanatory text boxes on the right. The diagram, which is the same as the first sub-image, depicts a stylized tree. At its apex is a radiant sun labeled \"Salvation (Moksha)\". The tree's trunk is labeled \"Philosophy\", and its roots are labeled with the words \"Artha\", \"Dharma\", \"Kama\", and \"Moksha\". The branches bear leaves with various philosophical school names, including \"Vedanta\", \"Yoga\", \"Nyaya\", \"Varuma\", \"Hravangu\", and others. To the right of this tree diagram, three rectangular text boxes with rounded corners and a subtle, iridescent background are arranged in a column. The top box states, \"Indian thought treats life as a layered journey, where reflection on existence guides choices and conduct.\" The middle box explains, \"The four aims of life—Artha, Dharma, Kama, Moksha—outline material prosperity, ethical duty, emotional fulfillment, and ultimate liberation.\" The bottom box concludes, \"Liberation (moksha) is envisioned as the summit where diverse paths of inquiry and practice finally converge.\" The overall aesthetic is clean and modern, using a harmonious color palette to present the information clearly.", "input_images": ["images/multi_image/747c5d7b-50a3-4101-a61e-27bbcb2485ba_input_0.jpg"], "target_image": "images/multi_image/747c5d7b-50a3-4101-a61e-27bbcb2485ba_target.png", "width": 1376, "height": 768}
+{"index": 240, "id": "c010b9ef-edd3-483b-a39c-6515805747d5", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This slide, titled \"CORE WORKFLOW FOCUS\" in large blue text at the top, is divided into four columns, each illustrating a key workflow area with an image and a bulleted list. The first input image is featured in the top-left corner under the header \"1. INDIVIDUAL PRODUCTIVITY,\" depicting a woman typing at a computer, accompanied by a list containing \"Prioritizing Daily Objectives,\" \"Executing Standard Activities,\" \"Tracking Personal Milestones,\" and \"Handling Routine Challenges.\" The second input image appears in the middle-left section under \"2. TEAM COMMUNICATION,\" showing a group meeting, with bullet points below reading \"Planning Huddles,\" \"Ongoing Status Sharing,\" \"Aligning Team Efforts,\" and \"Escalating Concerns Quickly.\" The third input image is located in the middle-right under \"3. DATA MANAGEMENT,\" showing a woman analyzing data on multiple screens, paired with the list \"Collecting Digital Records,\" \"Reviewing Key Indicators,\" \"Validating Information Quality,\" and \"Protecting Sensitive Files.\" Finally, the fourth input image is on the far right under \"4. TECHNICAL SUPPORT & UPKEEP,\" featuring a technician with tools, with the corresponding text \"Routine System Reviews,\" \"Optimizing Operational Settings,\" \"Caring for Critical Equipment,\" and \"Resolving Breakdowns Rapidly.\" The slide concludes with a footer at the bottom stating, \"SLIDE: DRIVING RELIABLE AND STREAMLINED WORKFLOWS.\"", "input_images": ["images/multi_image/c010b9ef-edd3-483b-a39c-6515805747d5_input_0.jpg", "images/multi_image/c010b9ef-edd3-483b-a39c-6515805747d5_input_1.jpg", "images/multi_image/c010b9ef-edd3-483b-a39c-6515805747d5_input_2.jpg", "images/multi_image/c010b9ef-edd3-483b-a39c-6515805747d5_input_3.jpg"], "target_image": "images/multi_image/c010b9ef-edd3-483b-a39c-6515805747d5_target.png", "width": 1376, "height": 768}
+{"index": 241, "id": "0140ba34-7511-4c14-8b21-7832ab4ce7fa", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Illustrating a colorful, educational infographic designed in a playful \"Labubu\" cartoon style, this slide explains the idea of absconding through three distinct vertical panels set against a clean white background. At the top, a bold title reads \"ALL ABOUT ABSCONDING,\" where \"ALL ABOUT\" is written in large, navy-blue, sans-serif capital letters, and \"ABSCONDING\" is in a matching large, bright orange font. To the left of the title, a small orange arrow points rightward, while to the right, small teal and purple dollar signs float among star-shaped sparkles.\n\nThe first panel on the left is titled \"BASIC IDEA OF 'ABSCOND'\" inside a blue rectangular header bar with white sans-serif text. Below this header is an illustration (corresponding to the first sub-image) showing a pink, furry, rabbit-eared creature with sharp teeth running away from a building labeled \"OFFICE\" with a red arrow pointing toward it. Beneath the image, a text block begins with the bold label \"MEANING:\" followed by the text: \"To go away in haste and in secret, usually to avoid being discovered, caught, or held responsible for wrongdoing, obligations, or debts.\"\n\nThe middle panel features an orange header bar with the white text \"TYPICAL EXAMPLES.\" This section contains three illustrations arranged vertically. The top illustration (matching the fourth sub-image conceptually, though visually distinct as a robber) depicts a person in business attire wearing a mask and sprinting off with bags of money and a document marked \"TAXES,\" accompanied by the text: \"Money-Related: Disappearing to dodge creditors, tax collection, or investigations into financial misconduct.\" The middle illustration (matching the second sub-image) shows the pink creature wearing handcuffs climbing a brick wall, with the text: \"Justice-Related: Failing to appear in court, escaping custody, or running off while under supervision.\" The bottom illustration (matching the third sub-image) depicts the creature leaving a reception desk with a suitcase while a woman waves \"Bye!\", and is paired with the text: \"Responsibility-Related: Quietly leaving a job, lease, or contract to avoid duties, payments, or promised work.\"\n\nThe third panel on the right has a red header bar with the white text \"POSSIBLE OUTCOMES.\" It features an illustration (matching the fourth sub-image) of a police officer chasing a running man with a net, while a gavel and chain lie in the foreground. Below this image, a dense block of text lists consequences using vertical bars as separators: \"Court-Ordered Warrants | New Criminal Charges (such as failure to appear) | Loss of Bail or Financial Guarantees | Harsher Sentences if Caught | Long-Term Damage to Legal and Personal Records.\" At the very bottom center of the slide, small grey text reads: \"Labubu Style Educational Series | Updated Edition.\"", "input_images": ["images/multi_image/0140ba34-7511-4c14-8b21-7832ab4ce7fa_input_0.jpg", "images/multi_image/0140ba34-7511-4c14-8b21-7832ab4ce7fa_input_1.jpg", "images/multi_image/0140ba34-7511-4c14-8b21-7832ab4ce7fa_input_2.jpg", "images/multi_image/0140ba34-7511-4c14-8b21-7832ab4ce7fa_input_3.jpg"], "target_image": "images/multi_image/0140ba34-7511-4c14-8b21-7832ab4ce7fa_target.png", "width": 1376, "height": 768}
+{"index": 242, "id": "c6654cbb-8ff2-4203-b734-c236898058f1", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### CREATIVE RENDERING BRIEF\nAct as a world-class presentation designer generating a single slide from structured JSON.\n\n### JSON SEMANTICS\n- `global_style`: the immutable design system for this slide deck. Apply it consistently and exactly.\n- `visual_hierarchy`: the intended reading order, emphasis, and atmosphere.\n- `source_images`: the images to embed in the slide.\n- `blocks`: the only renderable content elements.\n- `special_instructions`: explicit creative direction that takes priority when it exists.\n\n### NON-NEGOTIABLES\n- Do not hallucinate text, numbers, labels, logos, captions, or section names.\n- Keep all text within visible boundaries; adjust line breaks, scale, spacing, and weight to improve legibility.\n- Preserve referenced colors and fonts. When a block points to a palette color, use that exact color.\n- If the content is minimal, let the slide breathe. Do not fill empty space with invented graphics or copy.\n\n### COMPOSITION GUIDANCE\nCreate a visually distinct layout that suits the content density. Text-heavy slides can use structured columns or staggered cards; visual-heavy slides should give the visual element priority; mixed slides should balance imagery and text with clear hierarchy. Output a refined, high-fidelity slide image.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Model Comparison Results\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#F5F7FB\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #FFFFFF 0%, #E4EBF7 55%, #D9E2F2 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Montserrat\\\",\\n      \\\"secondary\\\": \\\"Open Sans\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#3961A8\\\",\\n      \\\"secondary\\\": \\\"#7E9AC9\\\",\\n      \\\"text\\\": \\\"#1F2933\\\",\\n      \\\"accent\\\": \\\"#22A779\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall look is clean and professional with a data-visualization feel. Soft blue tones complement the stacked bar chart colors, while subtle gradients give depth without distracting from the image. Corners are slightly rounded, and gentle drop shadows are used under the main chart container to make it feel elevated. No clutter; focus remains on the comparative results figure.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Analytical, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 21,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-21\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Tarsier2 Head-to-Head Evaluation\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#1F2933\\\",\\n            \\\"font_family\\\": \\\"Montserrat\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, above the chart with generous top margin\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Title is large and centered, spanning about 70% of the slide width, with tight letter spacing for a modern analytic feel.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/c6654cbb-8ff2-4203-b734-c236898058f1_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/c6654cbb-8ff2-4203-b734-c236898058f1_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Stacked bar chart comparing Tarsier2 performance against several alternative models.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, occupying roughly 75% of slide width and 60% of height\\\",\\n            \\\"image_style\\\": \\\"clean rendering with a thin border and soft drop shadow\\\",\\n            \\\"container\\\": \\\"rounded rectangle card with white background\\\",\\n            \\\"special_instructions\\\": \\\"Place the image inside a white card with 18px border radius, a 1px #D0D7E5 border, and a soft drop shadow (#000000 at 10% opacity, 16px blur). The card is centered horizontally and slightly above vertical center so the caption can sit beneath it.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-21-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Tarsier2 wins more often than baselines while maintaining a high rate of ties.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Open Sans\\\",\\n            \\\"color\\\": \\\"#394B63\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the chart card\\\",\\n            \\\"container\\\": null,\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Single line, medium weight, centered alignment; keep line length short so it sits comfortably under the chart without competing with it.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-21-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A large, soft gradient ring partially visible behind the chart card, fading from teal to blue to subtly emphasize the central image.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"centered behind the chart card, extending beyond its edges\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (about 0.25), blurred edges, no sharp lines; the ring should enhance focus on the chart without overpowering it.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate the final presentation slide image in widescreen 16:9 aspect ratio, with crisp typography and balanced composition.", "input_images": ["images/multi_image/c6654cbb-8ff2-4203-b734-c236898058f1_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 243, "id": "46c3c224-d761-40ff-984b-00d902a18bca", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide features a light yellow background with a dashed border and uses Minion characters to illustrate four stages of digital content handling. The title at the top reads \"UPLOADER: Central to Content Workflow,\" with \"UPLOADER\" in blue and the rest in yellow. The slide is divided into four vertical columns, each containing a corresponding sub-image and text.\n\nThe first column on the left, labeled \"CONTENT PREP,\" features the sub-image showing a Minion at a computer; the text below explains, \"Reviewing files before publishing. Converting items into supported file types. Grouping photos, clips, and documents for upload.\" The second column, \"UPLOAD PIPELINE,\" displays the sub-image of a Minion operating a control panel with a cloud icon; the description reads, \"Moving files from personal devices to the platform. Routing data through secure network channels. Tracking status indicators while content is sent.\" The third column, \"ACCESS & DISTRIBUTION,\" uses the sub-image of a Minion pointing at user icons; the text describes, \"Granting viewers entry to uploaded materials. Enabling teams to work together on shared files. Setting roles, permissions, and shareable links.\" The final column, \"ARCHIVE & RETENTION,\" shows the sub-image of a Minion next to stacked data boxes; the text notes, \"Placing content into structured libraries. Preserving older versions and backups. Periodically cleaning and reorganizing stored media.\" A footer runs along the bottom of the slide stating, \"A smart upload workflow powers smooth collaboration, reliable sharing, and long-term content care!\"", "input_images": ["images/multi_image/46c3c224-d761-40ff-984b-00d902a18bca_input_0.jpg", "images/multi_image/46c3c224-d761-40ff-984b-00d902a18bca_input_1.jpg", "images/multi_image/46c3c224-d761-40ff-984b-00d902a18bca_input_2.jpg", "images/multi_image/46c3c224-d761-40ff-984b-00d902a18bca_input_3.jpg"], "target_image": "images/multi_image/46c3c224-d761-40ff-984b-00d902a18bca_target.png", "width": 1376, "height": 768}
+{"index": 244, "id": "e4e6eb5b-c20f-4329-874b-29546b89d861", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "At the center, a serene nighttime slide titled **\"MOUNTAIN VILLAGE NIGHTS\"** in soft white letters is overlaid above the sub image of a brightly lit alpine town nestled in a snow-covered valley, with a sharp, iconic peak rising in the background beneath a clear, starry sky. The layout is divided into four subtle, translucent panels that frame the sub image without obscuring it.\n\nIn the **top-left** panel, a small heading **\"1. SETTING\"** is followed by brief text: *\"Snowy alpine valley • Quiet winter night • Clear horizon.\"*\n\nIn the **top-right** panel, under **\"2. SKY & PEAK\"**, short phrases read: *\"Deep blue night sky • Stars scattered above • Majestic mountain silhouette.\"*\n\nIn the **bottom-left** panel, labeled **\"3. VILLAGE LIGHTS\"**, minimal text notes: *\"Warm golden windows • Meandering streets • Homes clustered on slopes.\"*\n\nIn the **bottom-right** panel, titled **\"4. MOOD\"**, a few words summarize the atmosphere: *\"Calm • Cozy • Remote winter retreat.\"* The sub image remains the main visual focus, filling the background and drawing attention to the illuminated town and surrounding mountains.", "input_images": ["images/multi_image/e4e6eb5b-c20f-4329-874b-29546b89d861_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 245, "id": "490a027d-a87f-40d3-a59b-baaec356cc09", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "The slide features a white background with a horizontal layout organized into five distinct columns, topped by the main title \"PERFORMANCE MANAGEMENT IN LOGISTICS\" in large, bold, black text. The first column on the far left displays a sub-image of a white semi-truck, followed by icons of a checklist and a target; below these visuals is the bold heading \"STRATEGIC PRIORITIES\" and a bulleted list reading \"Clarify Outcomes\" (with sub-bullets \"Service reliability\" and \"Cost leadership & safety\"), followed by \"Define Success Measures\" (with sub-bullets \"Cost per shipment\", \"On‑time arrival ratio\", \"Vehicle fill rate\", \"Incident frequency\", and \"Carbon footprint per load\"). The second column features a sub-image of a desktop computer displaying a bar chart, accompanied by icons of a barcode scanner and cloud storage, with the bold heading \"INFORMATION CAPTURE\" and bullet points \"Operational Data Streams\" (sub-bullets \"Transport control systems\", \"Accounting & billing\", \"Maintenance logs\") and \"Market & Context Inputs\" (sub-bullets \"Customer feedback\", \"Open data sources\", \"Sector statistics\", \"Regulatory & trend analysis\"). The third column includes a sub-image of a laptop screen showing a graph, an icon comparing a person to a chart, the bold heading \"CROSS-UNIT REVIEW\", and the bullet points \"Assess results across depots, fleets, and service lines\" and \"Share winning practices and align internal standards\". The fourth column presents a sub-image of city buildings and trucks, the bold heading \"MARKET POSITIONING ANALYSIS\", and bullet points \"Benchmark against Market leaders\" (sub-bullet \"Direct and regional rivals\") and \"Learn from Exemplary operators\" (sub-bullet \"Innovative organizations in adjacent fields\"). Finally, the fifth column on the right displays a sub-image of a map with a gear and upward arrow, icons of a gear, arrow, and rising graph, the bold heading \"CONTINUOUS ENHANCEMENT LOOP\", and the bullet points \"Detect performance gaps and investigate drivers\", \"Design targeted initiatives and allocate resources\", and \"Roll out upgrades, review outcomes, and refine regularly\".", "input_images": ["images/multi_image/490a027d-a87f-40d3-a59b-baaec356cc09_input_0.jpg", "images/multi_image/490a027d-a87f-40d3-a59b-baaec356cc09_input_1.jpg", "images/multi_image/490a027d-a87f-40d3-a59b-baaec356cc09_input_2.jpg", "images/multi_image/490a027d-a87f-40d3-a59b-baaec356cc09_input_3.jpg"], "target_image": "images/multi_image/490a027d-a87f-40d3-a59b-baaec356cc09_target.png", "width": 1376, "height": 768}
+{"index": 246, "id": "8873dc1e-a527-433c-a157-529c45c01247", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are given a JSON description of one slide. Produce the final slide image with professional presentation quality.\n\nInterpretation rules:\n- `global_style` sets the exact design constraints and must be followed closely.\n- `blocks` contain all visible content. Do not create extra text.\n- `visual_hierarchy` tells you how to prioritize scale, position, contrast, and attention.\n- `special_instructions` should be treated as explicit art direction.\n- `source_images` are the images to embed in the slide.\n\nRendering rules:\n- Keep typography sharp, readable, and inside the canvas.\n- Use the provided colors and fonts consistently.\n- Prefer intentional whitespace over filler.\n- Adapt the layout to the material instead of using the same template every time.\n- Ensure the slide looks like part of a cohesive deck while still having its own visual identity.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Modern Educational Infographic\",\n    \"background\": {\n      \"color\": \"#1E1E1E\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#5D5FE8\",\n      \"secondary\": \"#FFFFFF\",\n      \"text\": \"#1A1A1A\",\n      \"accent\": \"#FFD700\"\n    },\n    \"special_instructions\": \"The slide uses a dark charcoal backdrop to highlight bright infographic cards. Maintain a playful, classroom-friendly look with rounded rectangles, soft shadows, and generous padding. Cards may use pastel pinks, teals, and creams to clearly separate concepts. Keep typography bold and simple, using strong headings and concise explanatory lines.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Educational, Playful, Informative\",\n    \"composition_style\": \"Grid-based, Three-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 6,\n      \"blocks\": [\n        {\n          \"id\": \"title-6\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Medicinal Syrups: Roles and Considerations\",\n          \"style\": {\n            \"color\": \"#5D5FE8\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"shadow\"\n              ],\n              \"outline_color\": null,\n              \"shadow_type\": \"drop\"\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Place the title in a white rounded banner with a thin purple border and a soft drop shadow. Center the text horizontally and give it enough padding to feel prominent but not cramped.\"\n          }\n        },\n        {\n          \"id\": \"body-6-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Key uses: sweeten bitter medicines, improve patient acceptance, and support dosing in children and older adults.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below title\",\n            \"container\": \"rounded panel with pink background\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Left-align the text within a light pink rounded rectangle. Use short line lengths and clear spacing for readability. Include a subtle pattern of lighter pink rounded shapes behind the text and a small informational icon in the top right corner of the panel.\"\n          }\n        },\n        {\n          \"id\": \"body-6-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Cautions: high sugar load, unsuitable for many diabetics, and potential for crystallization or spoilage if stored poorly.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"center column, below title\",\n            \"container\": \"rounded panel with light teal background\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Left-align the text inside a light teal rounded rectangle. Maintain even margins and line spacing. Add a subtle teal pattern to the background of the panel and place a small warning-style icon in the top right corner.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_912_260_1293_616.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_912_260_1293_616.jpg\"\n              }\n            ],\n            \"source\": \"Infographic illustrating syrup benefits and target patient groups\",\n            \"caption\": \"patient groups (ideal for children and elderly)\"\n          },\n          \"style\": {\n            \"position\": \"right column, below title\",\n            \"image_style\": \"contained within a cream-colored rounded panel\",\n            \"container\": \"rounded panel with cream background\",\n            \"special_instructions\": \"The image shows a central bottle of syrup with a spoon. Surrounding it are icons representing 'taste masking' (shield with mask), 'antioxidant effect' (shield with sun), a child, and an elderly person. Arrows connect these elements to the syrup bottle. The entire graphic is set against a cream-colored rounded rectangle.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nGenerate a refined 16:9 slide image suitable for a professional deck, with clear hierarchy, accurate content, and no unnecessary clutter.", "input_images": ["images/multi_image/8873dc1e-a527-433c-a157-529c45c01247_input_0.jpg"], "target_image": "images/multi_image/8873dc1e-a527-433c-a157-529c45c01247_target.png", "width": 1376, "height": 768}
+{"index": 247, "id": "4c86b07b-3fde-47a5-86c7-e71e6ef9b1ac", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Centered on a light blue background with a subtle dotted pattern, this presentation slide features the main title \"From Prep to Plate: Matching Tools to Tasks\" in bold, dark blue text at the top. Beneath the heading, two white text boxes introduce the theme, with the left box reading, \"Kitchen success starts with pairing each job to a tool designed for it,\" and the right box stating, \"The right choice boosts safety, speed, and the quality of every dish.\" The lower portion of the slide presents a horizontal row of five illustrations, each tagged with a different kitchen tool and its use. From left to right, these are: a chef slicing a baguette labeled \"Bread Knife for Clean Slices\"; hands using a round cutter on a sandwich labeled \"Shape Cutters for Decorative Bites\"; a person grating a carrot labeled \"Box Grater for Shredded Ingredients\"; a sandwich under a broiler labeled \"Overhead Broiler for Melting Toppings\"; and hands slicing a cabbage labeled \"Vegetable Chopper for Shredded Greens.\" The first sub-image shown is the chef slicing bread, the second shows the hands using a cookie cutter, the third depicts the grating of the carrot, and the fourth shows the sandwich under the salamander, all of which correspond to the first four illustrations in the final slide layout.", "input_images": ["images/multi_image/4c86b07b-3fde-47a5-86c7-e71e6ef9b1ac_input_0.jpg", "images/multi_image/4c86b07b-3fde-47a5-86c7-e71e6ef9b1ac_input_1.jpg", "images/multi_image/4c86b07b-3fde-47a5-86c7-e71e6ef9b1ac_input_2.jpg", "images/multi_image/4c86b07b-3fde-47a5-86c7-e71e6ef9b1ac_input_3.jpg"], "target_image": "images/multi_image/4c86b07b-3fde-47a5-86c7-e71e6ef9b1ac_target.png", "width": 1376, "height": 768}
+{"index": 248, "id": "d96a1930-ecb3-49d6-bc36-e17e00722aaf", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are an AI slide-generation engine. Transform the provided slide JSON into a polished, presentation-ready image.\n\nFollow these rules precisely:\n- Treat the JSON as the single source of truth. Render only the text contained in `blocks`; do not invent, paraphrase, or append any wording.\n- Embed every referenced source image into the layout.\n- Use `global_style` as the visual identity system; keep its typography, colors, background treatment, and overall design language intact.\n- Respect `visual_hierarchy` to determine dominant, secondary, and subtle elements.\n- Apply `special_instructions` wherever they appear; they can refine or override the general styling.\n\nAim for a sophisticated composition. Avoid defaulting to a simple title-and-bullets layout. Use asymmetry, modular sections, visual focal points, judicious white space, and varied structure (e.g., split views, cards, clusters) while maintaining clarity and legibility. The slide should look cohesive, modern, and ready for live presentation.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Dark Neon Gradient\",\n    \"background\": {\n      \"color\": \"#1A0B2E\",\n      \"gradient\": \"radial-gradient(circle at center, #2D1B4E 0%, #0F0518 100%)\",\n      \"texture\": \"none\",\n      \"pattern\": \"Abstract fluid shapes in corners (top-left blue/pink, top-right pink, bottom-right orange/blue)\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#FF0055\",\n      \"secondary\": \"#00D4FF\",\n      \"text\": \"#FFFFFF\",\n      \"accent\": \"#FFD700\"\n    },\n    \"special_instructions\": \"The slide features a dark, moody atmosphere with a deep purple/blue background. It uses a 'glassmorphism' aesthetic for content containers (semi-transparent rounded rectangles with subtle borders). The design incorporates neon glow effects, particularly on the title text and the infographic elements on the right. The overall mood is modern, urgent, and visually striking.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Modern, Urgent, Professional\",\n    \"composition_style\": \"Asymmetric, Split layout (Text left, Visual right)\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 2,\n      \"blocks\": [\n        {\n          \"id\": \"title-1\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Menorca Talayótica: Shared Responsibility, Shared Future\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"neon_glow\",\n                \"drop_shadow\"\n              ],\n              \"outline_color\": \"#FF0055\",\n              \"shadow_type\": \"drop\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, bold title with a vivid pink/red neon glow, slightly extending beyond the central text area for emphasis.\"\n          }\n        },\n        {\n          \"id\": \"body-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Ancient talayotic sites connect cultural memory, landscape, and community identity.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-top\",\n            \"container\": \"rounded rectangle panel, semi-transparent dark purple background\",\n            \"container_opacity\": \"0.7\",\n            \"special_instructions\": \"Place the text within a glassmorphism card. Add a minimal line icon of a stone monument to the left of the paragraph.\"\n          }\n        },\n        {\n          \"id\": \"body-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Inclusive heritage policies stimulate green jobs, local business, and social cohesion.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-bottom-left\",\n            \"container\": \"rounded rectangle panel, semi-transparent dark purple background\",\n            \"container_opacity\": \"0.7\",\n            \"special_instructions\": \"Text sits inside a glassmorphism card. A stylized green leaf icon appears to the right of the text, with a small group-of-people icon directly underneath.\"\n          }\n        },\n        {\n          \"id\": \"body-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Co-creation between institutions, residents, and visitors ensures resilient long-term stewardship.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"left-bottom-right\",\n            \"container\": \"rounded rectangle panel, semi-transparent dark purple background\",\n            \"container_opacity\": \"0.7\",\n            \"special_instructions\": \"Place the text in a glassmorphism card. Show a gear icon on the right and, beneath the text, a simplified network/flowchart icon to suggest collaboration and shared decision-making.\"\n          }\n        },\n        {\n          \"id\": \"image-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_848_244_1253_688.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_848_244_1253_688.jpg\"\n              }\n            ],\n            \"source\": \"Infographic showing the cycle of Menorca's heritage protection\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"right side, vertically centered\",\n            \"image_style\": \"none\",\n            \"container\": \"none\",\n            \"special_instructions\": \"The image shows a map of Menorca with a Talayotic monument, surrounded by a circular flow of icons (Euro symbol, people, gear, hands holding people) and two logos (UNESCO and Biosphere Reserve). The elements have a neon glow effect.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nRender the output as a single high-resolution 16:9 slide. Avoid introducing any additional text, decorative icons, or extraneous visual elements beyond what is specified.", "input_images": ["images/multi_image/d96a1930-ecb3-49d6-bc36-e17e00722aaf_input_0.jpg"], "target_image": "images/multi_image/d96a1930-ecb3-49d6-bc36-e17e00722aaf_target.png", "width": 1376, "height": 768}
+{"index": 249, "id": "6fdc56e4-6cf0-4cd0-85fa-27d91313e0a5", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### CREATIVE RENDERING BRIEF\nAct as a world-class presentation designer generating a single slide from structured JSON.\n\n### JSON SEMANTICS\n- `global_style`: the immutable design system for this slide deck. Apply it consistently and exactly.\n- `visual_hierarchy`: the intended reading order, emphasis, and atmosphere.\n- `source_images`: the images to embed in the slide.\n- `blocks`: the only renderable content elements.\n- `special_instructions`: explicit creative direction that takes priority when it exists.\n\n### NON-NEGOTIABLES\n- Do not hallucinate text, numbers, labels, logos, captions, or section names.\n- Keep all text within visible boundaries; adjust line breaks, scale, spacing, and weight to improve legibility.\n- Preserve referenced colors and fonts. When a block points to a palette color, use that exact color.\n- If the content is minimal, let the slide breathe. Do not fill empty space with invented graphics or copy.\n\n### COMPOSITION GUIDANCE\nCreate a visually distinct layout that suits the content density. Text-heavy slides can use structured columns or staggered cards; visual-heavy slides should give the visual element priority; mixed slides should balance imagery and text with clear hierarchy. Output a refined, high-fidelity slide image.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Model Distillation Performance Overview\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#0C1220\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #1C2640 0%, #050814 70%, #02030A 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": \\\"subtle faint grid lines suggesting a technical dashboard\\\"\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Montserrat\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#F2C94C\\\",\\n      \\\"secondary\\\": \\\"#2D9CDB\\\",\\n      \\\"text\\\": \\\"#FFFFFF\\\",\\n      \\\"accent\\\": \\\"#27AE60\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall look should feel like a clean, data-driven research slide. Dark navy background with a soft radial glow highlights the central chart. Use crisp, modern sans-serif fonts with high contrast white text. Minimal decorative elements; let the chart be the hero. Subtle grid or circuitry hints in the background can reinforce an analytical, machine-learning mood without distracting from the image.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Analytical, Modern\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 21,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-21\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Validation Perplexity Across Distillation Splits\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#FFFFFF\\\",\\n            \\\"font_family\\\": \\\"Montserrat\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with generous margin above the chart\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"All caps for key words 'VALIDATION PERPLEXITY'; slightly increased letter-spacing for a technical feel.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/6fdc56e4-6cf0-4cd0-85fa-27d91313e0a5_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/6fdc56e4-6cf0-4cd0-85fa-27d91313e0a5_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Line chart showing validation perplexity versus training steps for different teacher–student data splits.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center of the slide, spanning roughly 70% of width and 60% of height\\\",\\n            \\\"image_style\\\": \\\"subtle soft shadow and thin border\\\",\\n            \\\"container\\\": \\\"rounded rectangle with a very dark navy (#080C18) backdrop behind the image to separate it from the background\\\",\\n            \\\"special_instructions\\\": \\\"Apply a 4px border-radius to the image corners, a 1px border in #1F2937, and a soft, wide drop shadow to emphasize depth. Ensure the chart remains the dominant element.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-21-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Lower validation perplexity is achieved faster when more teacher data guides early training.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#A9B7D9\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the chart container\\\",\\n            \\\"container\\\": null,\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Small, centered text; keep to a single line to maintain a clean, research-figure style.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-21-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A faint, oversized circular gradient ring behind the chart, echoing the teal and green tones of the curves.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette with #2D9CDB and #27AE60 blended softly\\\",\\n            \\\"position\\\": \\\"centered behind the image block, extending slightly beyond its edges\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (~0.25), no hard edges, purely decorative to add depth without distracting from the chart.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nRender the slide as a high-quality 16:9 deck page. Use only the provided text, follow the global style, and make the composition feel intentional.", "input_images": ["images/multi_image/6fdc56e4-6cf0-4cd0-85fa-27d91313e0a5_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 250, "id": "ac2f7399-4ca3-4004-b648-7b27f366be69", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Split between tabular comparisons and concise metric summaries, this slide uses a light-colored background populated by simple geometric icons to explain evaluation results. The sub image with the numeric table appears centered in the slide, serving as the primary visual element. In the top-left, the slide features the title “COMPARING GENERATIVE MODELS” and in the top-center the subtitle “EVALUATION ACROSS SETTINGS.” Below the main title, minimal text labels the four columns as “CLEAN PADS,” “FULL,” “PADS,” and “PROMPT,” with a short note: “Lower is better.”\n\nDirectly beneath the sub image, a slim caption reads “PERFORMANCE SCORES PER MODEL.” To the right of the sub image, a compact callout box highlights: “BEST CLEAN PADS: SDXL” and “HIGHEST PADS: SD 3.” In the bottom-right corner, a brief takeaway states: “MODELS VARY BY CONDITION — CHOOSE SETTINGS TO MATCH YOUR GOAL.”", "input_images": ["images/multi_image/ac2f7399-4ca3-4004-b648-7b27f366be69_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 251, "id": "bb642acb-800d-4189-be34-7f39409ced50", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide adopts a clean, modern aesthetic set against a soft off‑white background with subtle warm lighting, evoking the calm atmosphere of a contemporary café interior. The layout is organized into three key zones: a concise header at the top, a compact pair of text blocks on the left, and a large visual focus area on the right featuring the sub image.\n\nAt the very top, a wide rectangular banner with gently rounded corners stretches across the slide. Its border is a thin, warm caramel line that echoes the tones of the leather chairs in the sub image. Centered within this banner, the title reads “Sunlit Café Corner” in a simple, bold sans‑serif font, rendered in a deep charcoal color. The text is horizontally aligned in the middle of the slide, with generous spacing above and below to maintain an airy feel.\n\nBeneath this header, the left half of the slide contains two compact text modules stacked vertically. Each module is enclosed within a minimal, rounded rectangle outlined in a light taupe shade. The corners are soft, mirroring the smooth curves of the café windows visible in the sub image.\n\n- The upper text box holds a single, short line: “Warm light filters across quiet, empty tables.” The font is a medium‑weight sans‑serif, aligned left, in a muted coffee‑brown color. The sentence is brief, inviting viewers to focus primarily on the visual mood.\n\n- Directly below, the second text box presents another minimal sentence: “Leather chairs and clean lines frame a peaceful interior.” This line uses the same font and color, maintaining consistency. The spacing between the two boxes is modest, keeping the text area compact and unobtrusive.\n\nThe right half of the slide is devoted almost entirely to the sub image, which is framed with a slender, light‑gray border and slightly rounded corners. The image is centered vertically, ensuring it becomes the dominant visual focal point. Within the sub image, rows of caramel and tan leather chairs face wooden tables, all bathed in golden sunlight that casts soft shadows across the surfaces. At the far side of the room, a person sits alone at a table, partially hidden behind a chair, emphasizing the quiet, contemplative ambiance. The back wall features a series of elongated rectangular windows, through which a lush outdoor scene is visible—green trees and blooming purple flowers—adding a contrasting burst of color to the warm interior tones.\n\nTo subtly anchor the visual, a faint drop shadow beneath the sub‑image frame gives the illusion that the photograph is gently lifted off the slide background, enhancing depth without distracting from the content. Overall, the slide composition prioritizes the sub image as the main storytelling element, with text kept to a minimum to simply hint at the calm, sunlit café setting depicted.", "input_images": ["images/multi_image/bb642acb-800d-4189-be34-7f39409ced50_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 252, "id": "f30e612a-ab58-449a-813b-823086e31857", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM\nYou are a cinematic presentation renderer with strong graphic design discipline.\n\n### OBJECTIVE\nGenerate a single polished slide image from the input JSON while preserving both semantic accuracy and deck-level style consistency.\n\n### RENDERING POLICY\n- Text fidelity is absolute: only render text from JSON `content` fields.\n- `global_style` controls the full visual language, including background, palette, fonts, texture, depth, lighting, and spacing.\n- `special_instructions` can introduce specific art direction such as atmosphere, materials, camera feel, or decorative treatment.\n- Maintain clear hierarchy: titles should lead, supporting text should be readable, captions should remain subtle, and visuals should not obscure essential text.\n- Use `source_images` to embed in the slide.\n\n### DESIGN DIRECTION\nChoose a layout that makes the slide memorable while staying coherent with the given style: bold hero composition, premium editorial layout, refined infographic, modular card layout, centered statement, or asymmetric grid. Keep the final output crisp, legible, balanced, and free of invented content.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Warm Lantern Festival Dessert\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#2B1A18\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #5C3A30 0%, #1A1010 70%)\\\",\\n      \\\"texture\\\": \\\"paper\\\",\\n      \\\"pattern\\\": \\\"subtle overlapping circles reminiscent of traditional window lattice, very low contrast\\\"\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Noto Serif SC\\\",\\n      \\\"secondary\\\": \\\"Noto Sans SC\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#E53935\\\",\\n      \\\"secondary\\\": \\\"#F5E1C8\\\",\\n      \\\"text\\\": \\\"#FDF6EE\\\",\\n      \\\"accent\\\": \\\"#F6A623\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Create an intimate, festive Chinese New Year ambiance that echoes the warm sunlight and wooden textures in the photo. Use deep maroon-brown background tones to let the off-white dessert bowl and bright red paper-cut details stand out. Typography should feel elegant and slightly traditional, with the title in a serif Chinese-style font and supporting text in a clean sans serif. Apply soft, warm lighting effects and gentle shadows beneath the central image to give a cozy tabletop feel. Incorporate subtle decorative motifs inspired by paper-cut patterns and round lantern shapes, keeping them low-contrast so the food image clearly dominates.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Warm, Festive, Elegant\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 26,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-26\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Sweet Reunion in a Bowl\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#FDF6EE\\\",\\n            \\\"font_family\\\": \\\"Noto Serif SC\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with comfortable margin above the image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Title in mixed case with slightly increased letter spacing; subtle golden drop shadow (#F6A623) to separate it from the dark background.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/f30e612a-ab58-449a-813b-823086e31857_input_0.png\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/f30e612a-ab58-449a-813b-823086e31857_input_0.png\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"A traditional bowl of tangyuan glutinous rice balls in sweet soup with goji berries, served in a warm wooden setting.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning about 70% of slide width and vertically centered between title and footer text\\\",\\n            \\\"image_style\\\": \\\"bordered with soft vignette shadow\\\",\\n            \\\"container\\\": \\\"rounded rectangle frame in deep brown (#3B2620) with thin inner border in warm beige (#F5E1C8)\\\",\\n            \\\"special_instructions\\\": \\\"Give the image slightly rounded corners (12–16px radius) and a soft, warm drop shadow to lift it off the background. Maintain original aspect ratio. The frame should be subtle so the photograph remains the dominant focus.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-26-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Tangyuan in ginger-infused sweet soup, shared under festival lanterns.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Noto Sans SC\\\",\\n            \\\"color\\\": \\\"#F5E1C8\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image frame\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Small, centered line of text; medium weight; keep it unobtrusive so it gently anchors the scene without drawing attention from the image.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-26-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two thin, semi-transparent circular motifs inspired by traditional red paper-cut patterns, placed partially off-canvas.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"decorative background shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"one large circle peeking from top-right corner, another smaller from bottom-left corner, both behind all other elements\\\",\\n            \\\"special_instructions\\\": \\\"Use #E53935 with 10–15% opacity so they appear as soft tonal decorations. Edges slightly blurred; ensure they never compete with the central image.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nRender the slide as a high-quality 16:9 deck page. Use only the provided text, follow the global style, and make the composition feel intentional.", "input_images": ["images/multi_image/f30e612a-ab58-449a-813b-823086e31857_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 253, "id": "7ace8a0d-a416-4e0e-a0eb-5ead8117239b", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide, titled \"UNLOCKING INSIGHT\" in large, bold, dark blue sans-serif capital letters centered at the top, features a clean, modern design against a pale blue background with faint geometric line patterns. The slide is structured into four vertical columns, each contained within a white rounded rectangle with a subtle drop shadow, presenting a progressive flow of ideas. At the very bottom, centered in smaller dark blue text, is the subtitle: \"A PATHWAY TO CURIOSITY AND GROWTH\".\n\nThe first column on the left is headed by the icon from the first sub-image, depicting a stylized human head in profile facing left, rendered in peach tones with a blue question mark inside and a blue gear behind it, surrounded by yellow sparkles. Below this icon, the bold black header reads \"CLARIFYING CURIOSITY\". The text beneath it, in a standard black sans-serif font, states: \"A mindset of questioning, exploration, and focus. Triggered by puzzling problems, fresh perspectives, or unfamiliar ideas.\"\n\nThe second column features the icon from the second sub-image, showing a blue and white telescope on a tripod pointing towards yellow stars, alongside a compass with a golden rim. The bold black header below reads \"CHANNELS OF EXPLORATION\". The descriptive text lists: \"Natural worlds (ecosystems, skies, oceans) Creative expression and design Scientific inquiry and technology Everyday experiences and human stories\".\n\nThe third column displays the icon from the third sub-image, which consists of a bar chart with ascending yellow and blue bars, overlaid by a rising blue arrow, topped by a glowing yellow lightbulb. The bold black header is \"OUTCOMES OF CURIOSITY\". The text below details: \"Strengthens problem‑solving and innovation. Builds confidence and resilience. Deepens understanding and perspective. Fuels ongoing learning.\"\n\nThe final column on the right utilizes the icon from the fourth sub-image, showing a hand in a blue sleeve holding soil with a green sprout growing from it, surrounded by yellow sparkles. The bold black header reads \"NURTURING GROWTH\". The accompanying text advises: \"Set aside time for reflection and inquiry. Engage in experiments and small trials. Seek diverse voices and new environments. Embrace uncertainty as a learning space. Turn insights into everyday practice.\"", "input_images": ["images/multi_image/7ace8a0d-a416-4e0e-a0eb-5ead8117239b_input_0.jpg", "images/multi_image/7ace8a0d-a416-4e0e-a0eb-5ead8117239b_input_1.jpg", "images/multi_image/7ace8a0d-a416-4e0e-a0eb-5ead8117239b_input_2.jpg", "images/multi_image/7ace8a0d-a416-4e0e-a0eb-5ead8117239b_input_3.jpg"], "target_image": "images/multi_image/7ace8a0d-a416-4e0e-a0eb-5ead8117239b_target.png", "width": 1376, "height": 768}
+{"index": 254, "id": "a1807b43-103b-400d-90b2-185c6264984f", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Under the heading **“CLIP REFERENCE: REPRESENTATION COMPARISON!”**, this comic-style slide highlights how different representations affect performance using cute characters.  \nThe sub image is placed prominently in the center of the slide, framed like a control dashboard.\n\nOn the **top-left**, a small character points to the “Full” row saying:  \n**“FULL DATA: BEST MATCH!”**\n\nOn the **bottom-left**, another character circles the “Prompt” and “Pads” rows with the caption:  \n**“PROMPTS & PADS: STILL STRONG!”**\n\nTo the **right** of the sub image, three simple callout bubbles explain the trend:\n\n1. **“MORE CONTEXT → HIGHER IMAGE SCORE”**  \n2. **“PROMPT SCORES STAY LOWER THAN IMAGE”**  \n3. **“CLEAN DATA: SIMPLE BUT WEAKEST”**\n\nAt the **bottom-center**, a tiny note reads:  \n**“COMPARE IMAGE vs PROMPT COLUMNS TO SEE THE DROP-OFF.”**", "input_images": ["images/multi_image/a1807b43-103b-400d-90b2-185c6264984f_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 255, "id": "558607d5-b5b2-4b8f-8791-65fa7415d0a7", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Organized into a four-quadrant layout against a blue background, this slide titled \"Cloud Strategy & Integrated Governance\" in large white text at the top presents a structured model for enterprise cloud adoption. The first input image of a cloud texture appears in the bottom-left, while the second input image of a balance scale is centered above the second section, and the third input image showing a network of people around a building is positioned in the top-right corner. The content is divided into four numbered points: \"01 VISION & ALIGNMENT: Define a clear cloud mission that supports business priorities and stakeholder expectations. (Placeholder text: Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.)\" is on the left; \"02 COST & VALUE MANAGEMENT: Balance financial efficiency with performance and innovation in the cloud environment. (Placeholder text: Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.)\" sits in the center-left; \"03 GOVERNANCE & COMPLIANCE: Coordinate policies, risk controls, and stakeholder roles across cloud platforms. (Placeholder text: Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.)\" is in the top-right area; and \"04 OPERATIONS & CONTINUOUS IMPROVEMENT: Optimize reliability, security, and collaboration through iterative enhancement. (Placeholder text: Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.)\" is located in the bottom-right. Each section includes a colored circle with the number (red, yellow, green, blue respectively) and a block of placeholder text.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/558607d5-b5b2-4b8f-8791-65fa7415d0a7_input_0.jpg", "images/multi_image/558607d5-b5b2-4b8f-8791-65fa7415d0a7_input_1.jpg", "images/multi_image/558607d5-b5b2-4b8f-8791-65fa7415d0a7_input_2.jpg"], "target_image": "images/multi_image/558607d5-b5b2-4b8f-8791-65fa7415d0a7_target.png", "width": 1376, "height": 768}
+{"index": 256, "id": "b900fe21-c8c4-453e-a515-19b18dc74661", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are an elite AI Slide Composition Engine. Your capabilities fuse the visual instincts of a senior Creative Director with the systematic rigor of an Information Architect.\n\n**Primary Mission:** Transform the incoming **Structured Content JSON** into a single, production-ready presentation slide image.\n\n### INPUT DATA INTERPRETATION\nA JSON object describing ONE slide will be provided. Interpret its keys as follows:\n1. **`global_style`** – Establishes the non-negotiable visual universe for the slide (background, lighting, master palette, typography rules).\n2. **`visual_hierarchy`** – Guides layout balance, emphasis, and overall narrative feel.\n3. **`blocks`** – Enumerates each renderable element (text, images, icons, diagrams, shapes).\n\n### CORE EXECUTION DOCTRINES (MANDATORY)\n\n#### 1. IMAGE FIDELITY & PRESERVATION (TOP PRIORITY)\n* **Directive:** For every block with `type: \"image\"`, you MUST place the referenced source image(s) directly into the slide.\n* **Source Assets:** All images are provided as attachments with this prompt and must be treated as final art.\n* **Hard Restrictions:**\n  * **DO NOT** crop, rescale in a way that distorts, repaint, or reinterpret any embedded chart, diagram, or picture.\n  * **DO NOT** swap these with stock images or AI-generated substitutes.\n  * **DO NOT** omit or obscure them; each referenced image must be visible in the final frame.\n* **Permitted Enhancements:** You MAY stylize the image frame or wrapper (e.g., add subtle depth, border, glow, or tilt) as long as all informational content remains crisp and legible.\n* **Multiple Images Handling:** When more than one image is specified, lay them out intelligently (grid, row, column, or as directed by `special_instructions`).\n\n#### 2. TEXT USAGE & LIMITS\n* **Directive:** The JSON `content` fields are the ONLY allowed textual source.\n* **Strict Limits:**\n  * **No Invented Copy:** Do not create new headlines, labels, or supporting text beyond what appears in the JSON.\n  * **No Filler:** If the slide feels open or minimal, preserve that minimalism—use whitespace as a deliberate design choice.\n  * **Formatting Flexibility:** You may tune line breaks, size, and emphasis (weights, small caps, spacing) to enhance clarity, but must avoid dense, unreadable text blocks.\n* **`role: \"script\"` Treatment:** Any block labeled `script` is informational for you only. It shapes tone and style but must remain invisible unless `special_instructions` explicitly says to render it.\n\n#### 3. SPECIAL_INSTRUCTIONS PRIORITY CHANNEL\n* **Directive:** Any `special_instructions` field (whether under `global_style` or an individual block) functions as the creative override layer.\n* **Behavior:** When present, its guidance supersedes general defaults.\n  * Example: If `global_style` suggests “neutral studio background” yet `special_instructions` requests “high-tech holographic desk,” adopt the holographic look.\n  * Use this channel to justify additional lighting, texture, micro-animations-in-spirit (conveyed visually), or other stylistic refinements.\n\n#### 4. GLOBAL_STYLE ENFORCEMENT (RIGID CONSISTENCY)\n* **Directive:** Treat `global_style` as your master design system.\n* **Background:** Apply the exact `background.color` and any gradients/patterns without deviation across all slides generated from this style.\n* **Typography:**\n  * Titles must use `font_palette.primary`.\n  * Body and supporting text must use `font_palette.secondary`.\n* **Color System:** Whenever a block references a `color_ref`, map it precisely to the specified hex in `color_palette`.\n* **Role-Based Sizing & Placement:**\n  * All TITLE elements (`importance: \"dominant\"` or equivalent) must share consistent scale and anchor position.\n  * Body/list elements marked as subheadings or core points must share uniform styling from slide to slide.\n  * Caption-level text must likewise be consistent in size, weight, and alignment.\n* **Intent:** Independent slides rendered from the same `global_style` should appear as a coherent deck with identical typography rhythm and chromatic identity.\n\n#### 5. LAYOUT VARIETY & COMPOSITIONAL INTEREST\n* **Directive:** Avoid generic “centered title with stacked bullets” layouts unless explicitly required.\n* **Implementation Guidance:**\n  * **Text-centric slides:** Consider multi-column layouts, asymmetric grids, or offset groupings rather than a single vertical stack.\n  * **Visual-first slides:** Allow diagrams, illustrations, and photos to dominate; keep text compact, serving as labels or concise explanations.\n  * **Hybrid slides:** Design balanced compositions—side-by-side clusters, overlapping layers, or modular cards.\n  * **Whitespace Discipline:** Do not artificially enlarge text or add fictitious elements to consume space. Negative space should feel deliberate and premium.\n* **Variation Techniques:**\n  * Shift alignments (left, right, centered, or justified blocks).\n  * Vary container shapes (cards, bands, callouts) while respecting `global_style`.\n  * Adjust how visual weight is distributed (some slides highly image-driven, others typographically led).\n* **Outcome:** Each slide should have a distinct composition but still clearly belong to the same design system.\n\n### OUTPUT SPECIFICATION\nRender one high-resolution 16:9 slide image that complies with all style rules and faithfully represents the JSON structure.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Technical Engineering Blueprint\",\n    \"background\": {\n      \"color\": \"#F5F7FA\",\n      \"gradient\": \"radial-gradient(circle, #FFFFFF 0%, #E8EDF2 100%)\",\n      \"texture\": \"none\",\n      \"pattern\": \"none\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#6B9EC5\",\n      \"secondary\": \"#D4C5A9\",\n      \"text\": \"#2C3E50\",\n      \"accent\": \"#5D7996\"\n    },\n    \"special_instructions\": \"Maintain a precise, engineering-oriented visual language over a soft radial gradient background. Highlight a central 3D truss bridge rendered with realistic metallic lighting and shadows. Use cool steel blues with warm beige accents on structural parts. Titles should appear in bold, block-style lettering; descriptive copy must be set in a clean sans-serif. Organize content with one large hero visual above and four clearly separated informational columns in the lower portion of the slide.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Analytical, Instructional, Professional\",\n    \"composition_style\": \"Top-Aligned Title with Hero Visual and Quad-Column Footer Layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 818,\n      \"blocks\": [\n        {\n          \"id\": \"title-818\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"STEEL TRUSS STRUCTURES\",\n          \"style\": {\n            \"color\": \"#6B9EC5\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"3D extrusion\",\n                \"drop shadow\"\n              ],\n              \"outline_color\": \"#FFFFFF\",\n              \"shadow_type\": \"drop\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Render the title as large 3D-styled lettering in light steel blue with a white edge highlight and a subtle bevel, casting a soft shadow onto the background.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_81_143_1300_492.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_81_143_1300_492.jpg\"\n              }\n            ],\n            \"source\": \"Main diagram of a steel truss bridge with labeled components\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-center, spanning full width below title\",\n            \"image_style\": \"3D rendered illustration with realistic lighting and shadows\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Use the provided image of the blue and beige truss bridge resting on supports, preserving all callouts for 'TOP CHORD', 'BOTTOM CHORD', 'DIAGONAL MEMBERS', 'VERTICAL MEMBERS', and 'GUSSET PLATE JOINT'.\"\n          }\n        },\n        {\n          \"id\": \"body-818-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"DEFINITION\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"dominant\",\n            \"position\": \"bottom-left, column header\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Serve as the bold heading for the first information column.\"\n          }\n        },\n        {\n          \"id\": \"illustration-818-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A small 2D schematic diagram of a truss structure with arrows indicating forces at the joints\",\n          \"style\": {\n            \"illustration_style\": \"technical diagram\",\n            \"color_scheme\": \"blue and black lines\",\n            \"position\": \"below DEFINITION header\",\n            \"special_instructions\": \"Create a minimalist line drawing of a truss with force arrows applied at the joints.\"\n          }\n        },\n        {\n          \"id\": \"body-818-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"A structure of slender members (chords & web) joined at their endpoints, designed to support loads. Forms stable triangular units.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left, below diagram\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Left-align this explanatory paragraph within the first column.\"\n          }\n        },\n        {\n          \"id\": \"body-818-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"KEY PRINCIPLES\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"dominant\",\n            \"position\": \"bottom-left-center, column header\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold heading for the second column of information.\"\n          }\n        },\n        {\n          \"id\": \"illustration-818-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A diagram showing two horizontal beams with arrows indicating tension and compression forces\",\n          \"style\": {\n            \"illustration_style\": \"technical diagram\",\n            \"color_scheme\": \"blue and beige beams with black arrows\",\n            \"position\": \"below KEY PRINCIPLES header\",\n            \"special_instructions\": \"Depict two straight members with opposing arrows clearly labeled as tension and compression.\"\n          }\n        },\n        {\n          \"id\": \"body-818-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Primarily axial force members.\\n• Experience tensile (TENSION) or compressive (COMPRESSION) forces.\\n• Little to no bending moment within members.\\n• Joint loads are assumed.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left-center, below diagram\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Display as a concise bulleted list, keeping the key force terms in parentheses exactly as written.\"\n          }\n        },\n        {\n          \"id\": \"body-818-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"MAJOR COMPONENTS\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"dominant\",\n            \"position\": \"bottom-right-center, column header\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold header for the third information column.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_724_556_1002_624.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_724_556_1002_624.jpg\"\n              }\n            ],\n            \"source\": \"Diagram showing truss joints\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"below MAJOR COMPONENTS header\",\n            \"image_style\": \"2D schematic\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Insert the given truss schematic, ensuring the 'JOINTS' label remains clearly legible.\"\n          }\n        },\n        {\n          \"id\": \"body-818-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"CHORDS: Top & bottom members, bear main loads.\\nWEB MEMBERS: Diagonal & vertical members, resist shear forces.\\nJOINTS (NODES): Connection points for members.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right-center, below diagram\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Format as a short list of component descriptions; emphasize the term before each colon through weight or small caps.\"\n          }\n        },\n        {\n          \"id\": \"body-818-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"ADVANTAGES & USES\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"dominant\",\n            \"position\": \"bottom-right, column header\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold heading text for the fourth column.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1035_556_1315_625.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1035_556_1315_625.jpg\"\n              }\n            ],\n            \"source\": \"Icons representing bridges, roofs, and cranes\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"below ADVANTAGES & USES header\",\n            \"image_style\": \"flat icons\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Place the provided trio of line icons—bridge, roofed building, and crane—centered under the header.\"\n          }\n        },\n        {\n          \"id\": \"body-818-8\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Efficient and lightweight.\\n• Long-span capabilities.\\n• High stiffness and strength-to-weight ratio.\\n• Cost-effective for large spans.\\nUsed in: BRIDGES, ROOFS FOR STADIUMS & HALLS, CRANES, INDUSTRIAL BUILDINGS.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right, below icons\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Render a compact bullet list of advantages followed by the application line, maintaining the capitalization exactly as supplied.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nGenerate a polished 16:9 slide image with precise spacing, clear hierarchy, and unwavering adherence to the JSON-defined rules and assets.", "input_images": ["images/multi_image/b900fe21-c8c4-453e-a515-19b18dc74661_input_0.jpg", "images/multi_image/b900fe21-c8c4-453e-a515-19b18dc74661_input_1.jpg", "images/multi_image/b900fe21-c8c4-453e-a515-19b18dc74661_input_2.jpg", "images/multi_image/b900fe21-c8c4-453e-a515-19b18dc74661_input_3.jpg"], "target_image": "images/multi_image/b900fe21-c8c4-453e-a515-19b18dc74661_target.png", "width": 1376, "height": 768}
+{"index": 257, "id": "92bb6053-5b0e-4da6-b9cb-37acc62920aa", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide features a clean white background with a single‑column, research‑oriented layout illustrating blockchain security analysis using a schematic diagram. At the very top, centered in large, bold, black, uppercase sans‑serif text, is the title: **\"CROSS-CHAIN EXECUTION TRACE ANALYSIS\"**. Directly beneath this, a slightly smaller subtitle in the same bold black font reads: **\"GRAPH-BASED DETECTION OF CROSS-CHAIN ATTACKS\"**.\n\nThe main visual content is dominated by the sub image placed centrally in the slide, occupying most of the vertical space. The sub image is a monochrome flow diagram framed by a thin black outline, visually guiding the viewer from top to bottom through the analysis pipeline:\n\n- At the very top of the sub image, a rounded rectangle labeled **\"Transaction Hash\"** connects via a right-pointing arrow to another rounded rectangle labeled **\"Traces and Logs\"**, indicating the input data sources.\n- Below these, a large rectangular block spans the width of the sub image, labeled **\"Cross-chain Execution Trace Graph Construction\"**. Inside this block, stylized node-and-arrow icons depict smart contract interactions. Circular nodes represent **addresses and functions**, while square nodes depict **log events**. Arrows labeled **\"CALL\"** and **\"CREATE\"** emphasize different types of contract interactions. On the right side of this block, a dashed box contains a simple legend explaining the circle and square symbols.\n- From the bottom of this construction block, two arrows diverge symmetrically downward into two smaller rectangles placed side by side. The left rectangle is titled **\"Global Graph Mining\"** and includes a small network icon. The right rectangle is titled **\"Local Graph Mining\"** and features a bar-chart icon, together suggesting complementary analytic perspectives on the graph structure.\n- Below these, arrows converge into a circular plus sign that feeds into a wider rectangle labeled **\"Supervised Multiple Classifiers\"**, indicating that features from both global and local graph mining are combined for machine-learning‑based decision making.\n- Finally, an arrow points downward to the last wide rectangle at the bottom labeled **\"Cross-chain Attacks / Not Attack\"**, representing the binary classification outcome.\n\nBeneath the sub image, aligned centrally and set in small, light‑weight black sans‑serif font, is a concise caption: **\"From transaction traces to graph mining and supervised classification for cross-chain attack detection.\"** No additional decorative elements are used, maintaining a minimal, technical aesthetic that keeps the sub image as the primary focus of the slide.", "input_images": ["images/multi_image/92bb6053-5b0e-4da6-b9cb-37acc62920aa_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 258, "id": "1ccadd37-ab95-40f2-8a5a-32af5ccaae98", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide, titled **\"LANTERNS IN THE SKY,\"** adopts a clean, poster-like aesthetic that emphasizes vertical space and an upward gaze. The background is a muted, off‑white canvas framed by a thin, dark border with slightly roughened edges, giving a subtle, printed-on-paper feel. Along the very top edge, the main title **\"LANTERNS IN THE SKY\"** appears in a large, bold, serif typeface. The words are set in all caps, with “IN THE” slightly smaller and nestled between “LANTERNS” and “SKY.” A single, thin horizontal rule sits directly beneath the title, visually separating it from the rest of the layout. Below this line, in a small, letterspaced serif font, is the understated subtitle: **\"LIGHT, COLOR, CELEBRATION\"**.\n\nDominating the central portion of the slide is the **sub image**, placed prominently and aligned so that it becomes the visual anchor of the entire composition. The sub image shows a view looking straight up between two traditional buildings, revealing rows of vivid red lanterns hanging overhead against a bright blue sky scattered with soft white clouds. The lanterns are strung in parallel lines that run horizontally across the frame, creating a rhythmic pattern of circular shapes. Each lantern’s rich red tone contrasts sharply with the cool blue of the sky, while the dark rooflines and architectural details of the surrounding buildings form a natural frame around the open space above. The perspective makes the viewer feel as though they are standing at street level and gazing directly upward into a corridor of color and light.\n\nThe sub image is enclosed within a simple, thin, black rectangular outline that subtly separates it from the background while preserving the image’s vibrancy. No additional graphics or icons overlap the picture; the design intent is to allow the photograph’s composition—its symmetry, color contrast, and sense of depth—to carry the emotional weight of the slide.\n\nBeneath the sub image, occupying the bottom third of the slide, is a compact, centered text block designed to echo the structure of a vintage caption or placard. The block is framed with a light decorative border featuring minimalist corner flourishes that hint at traditional scrollwork without becoming ornate. Inside this box, the primary statement is set in a bold, all‑caps serif font:\n\n**\"STREETS OF LIGHT, SKY OF HOPE.\"**\n\nDirectly under this main line, in smaller, widely spaced capital letters, are three brief, separated phrases that serve as quiet thematic prompts rather than detailed explanations. Each phrase is divided by a small, solid circle:\n\n**\"SHARED CELEBRATION · FLOATING WISHES · COLORS ABOVE US\"**\n\nThe overall layout keeps text to a minimum, ensuring that the viewer’s attention remains firmly on the sub image, while the chosen words gently guide the interpretation: a moment of festivity, community, and upward-looking optimism. The monochrome framing elements and restrained typography ensure that the intense reds and blues of the lantern-filled sky remain the slide’s central and most memorable feature.", "input_images": ["images/multi_image/1ccadd37-ab95-40f2-8a5a-32af5ccaae98_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 259, "id": "49ca8830-e57f-4422-8160-ae279c91b17c", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are an AI slide-generation engine. Convert the supplied slide JSON into a high-quality presentation image.\n\nFollow these constraints exactly:\n- The JSON is authoritative. Render only the contents of `blocks`; never add new words.\n- Embed all source images into the slide.\n- `global_style` defines the shared visual identity. Preserve its fonts, colors, background, and style language.\n- `visual_hierarchy` describes what should feel dominant, secondary, or subtle.\n- `special_instructions` may refine or override the default look; respect it when present.\n\nDesign with variety. Do not default to a title plus bullet list. Use asymmetry, columns, modular cards, large focal visuals, restrained annotations, or generous negative space when appropriate. The final slide must feel clean, readable, and presentation-ready.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Ablation Study Metrics Focus\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#0F172A\\\",\\n      \\\"gradient\\\": \\\"linear-gradient(135deg, #020617 0%, #111827 45%, #1E293B 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": \\\"Subtle, semi-transparent diagonal grid lines fading toward the center\\\"\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Helvetica Neue\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#E5E7EB\\\",\\n      \\\"secondary\\\": \\\"#9CA3AF\\\",\\n      \\\"text\\\": \\\"#F9FAFB\\\",\\n      \\\"accent\\\": \\\"#38BDF8\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall feel is technical and clean, suitable for a research presentation. The dark navy gradient background makes the white table stand out sharply. Minimal decorations keep attention on the metrics image. Use crisp, sans-serif typography with clear spacing. Subtle cyan accent lines echo the table’s grid while not overpowering it.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Technical, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 22,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-22\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Ablation of Control Modules and Renderer\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#F9FAFB\\\",\\n            \\\"font_family\\\": \\\"Helvetica Neue\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center with generous margin from edges\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"drop-shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Single line, large size, centered. Apply a soft shadow to separate the text from the dark gradient.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-22-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Full configuration (w/ all) yields the best ISR, MIOU, and SR performance.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#E5E7EB\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"centered below title with moderate spacing\\\",\\n            \\\"container\\\": null,\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Small, single-line subtitle, horizontally centered and aligned with the title.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/49ca8830-e57f-4422-8160-ae279c91b17c_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/49ca8830-e57f-4422-8160-ae279c91b17c_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Table summarizing ablation results for different control components and a detail renderer across ISR, MIOU, and SR metrics.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center of slide, spanning about 70% of slide width\\\",\\n            \\\"image_style\\\": \\\"no border; subtle outer glow\\\",\\n            \\\"container\\\": \\\"rounded rectangle with very dark slate fill (#020617) and thin cyan outline\\\",\\n            \\\"special_instructions\\\": \\\"The image should be the dominant element, larger than all text. Place it in a slightly elevated card with 12px rounded corners and a soft cyan outer glow to emphasize importance.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-22-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two faint cyan horizontal lines extending left and right from the sides of the image card, suggesting alignment with the metrics table.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"decorative background shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"behind the image, aligned with its vertical center\\\",\\n            \\\"special_instructions\\\": \\\"Lines should be thin, semi-transparent, and extend toward the slide edges without touching them, sitting behind the image to subtly frame it.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate a single widescreen slide from the JSON. The output should be refined, readable, and consistent with the specified visual system.", "input_images": ["images/multi_image/49ca8830-e57f-4422-8160-ae279c91b17c_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 260, "id": "4c4ea09b-7909-4e50-b57b-1c1d35e0ce81", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Set against a soft cream background with subtle blue floral motifs in each corner, the slide introduces a title in graceful blue cursive script reading **“Driving Growth: A Story in Motion.”** Beneath, a single line in light serif text notes, **“A focused strategist sharing the vision, as data and global connections come to life on screen.”**  \n\nAt the center-right of the slide, the sub image is prominently placed, showing the presenter gesturing toward a glowing world map and rising charts behind her. To the lower-left, a small caption in muted blue reads, **“Data, people, and markets aligned toward sustainable growth.”** The remaining space is kept open and uncluttered, allowing the sub image to remain the primary focal point.", "input_images": ["images/multi_image/4c4ea09b-7909-4e50-b57b-1c1d35e0ce81_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 261, "id": "881e296b-4b0d-4ba4-b20c-ae2de22f77b4", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The slide features a white background with a clean, informative layout combining cartoon illustrations and text sections. At the very top, centered, is the title \"UNDERSTANDING CIVIL LIBERTIES\" written in a playful, hand-drawn style font with a teal fill and a darker green outline.\n\nOn the left side of the slide, the \"sub-image corresponding to the cartoon characters\" is positioned. This illustration depicts two characters: a shorter boy in a yellow shirt looking concerned, and a taller man in a white lab coat gesturing wildly. Above the boy are two thought bubbles: the top one contains a megaphone and a fist with the word \"OPPOSITION\" written inside, and the bottom one shows a jail cell. The taller character has a speech bubble pointing to him that reads: \"• Rights that protect people from state abuse! • Freedom to speak, protest, and organize. • Remember, Morty, power needs limits!\"\n\nTo the right of the cartoon characters, the slide is divided into two main text sections. The upper right section is headed by a jagged, cloud-like blue banner containing the text \"CORE FREEDOMS\" in a stylized font. Below this banner, three bullet points are listed in black text: \"• Freedom of expression and the press\", \"• Freedom of peaceful assembly and protest\", and \"• Protection from arbitrary arrest and detention\". To the immediate right of this text is the \"sub-image showing prisoners behind bars,\" where several cartoon figures of diverse appearances hold signs reading \"FREE SPEECH\" and \"DISSENT\".\n\nBelow the \"CORE FREEDOMS\" section is another text block headed \"WHY THEY MATTER\" in a font matching the \"CORE FREEDOMS\" style. This section lists three bullet points: \"• Allow citizens to challenge injustice\", \"• Prevent concentration of unchecked power\", and \"• Form the foundation of a healthy democracy\".\n\nAt the bottom right corner of the slide, there is a small circular icon depicting a globe with a padlock over it. Next to this icon, the text \"WORLDWIDE RESPONSIBILITY – DEFEND RIGHTS\" is written in a simple, dark sans-serif font. The overall color palette consists of white, teal, green, yellow, and black, creating a clear and educational visual presentation.", "input_images": ["images/multi_image/881e296b-4b0d-4ba4-b20c-ae2de22f77b4_input_0.jpg", "images/multi_image/881e296b-4b0d-4ba4-b20c-ae2de22f77b4_input_1.jpg"], "target_image": "images/multi_image/881e296b-4b0d-4ba4-b20c-ae2de22f77b4_target.png", "width": 1376, "height": 768}
+{"index": 262, "id": "7e626483-9105-4aeb-84a3-a16f4885d8e9", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide features a clean, light cream background with a subtle, desaturated topographic contour pattern overlaid. The slide is titled **\"TERRACED RICE LANDSCAPES: PATTERNS OF RURAL LIFE\"** in large, bold, dark brown sans‑serif font at the top center, underlined by a thick horizontal sepia line.\n\nOn the right side of the slide, a large rounded‑corner rectangular frame prominently displays the provided **sub image**. The sub image shows layered rice terraces stretching across a hillside, dotted with haystacks and small groups of people working the fields. Scattered village houses sit above the terraces, partially hidden by trees, giving a sense of depth and human presence in the landscape. The photo’s muted, almost monochrome color grading emphasizes texture, light, and the repeating geometry of the terraces rather than bright color.\n\nTo the left of the slide, a concise bulleted list in dark brown text, each bullet marked by a solid sepia diamond, offers minimal context so the image remains central:\n\n- Terraced fields tracing the contours of the hills  \n- Harvested rice bundled into rhythmic haystacks  \n- Villagers moving through irrigation paths and paddies  \n- Homes overlooking the fields, linking work and dwelling  \n\nBeneath the bullet list, a short caption line in smaller italic serif font reads:\n\n*Sub image: layered topography of agriculture, labor, and settlement captured in a single frame.*\n\nThe overall color palette uses warm browns, soft creams, and muted greens to complement the tones of the sub image, maintaining a calm, documentary aesthetic that keeps visual attention on the photograph as the primary content of the slide.", "input_images": ["images/multi_image/7e626483-9105-4aeb-84a3-a16f4885d8e9_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 263, "id": "ea664fbf-ab62-40f2-ad8e-298ef344d873", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This infographic slide, titled **\"Heroic Vigilance (The Caped Guardian)\"** in large bold text at the top, showcases a minimal, clean layout where the sub image is the central focus on a white background. The middle of the slide prominently features the sub image: a classic, comic-style caped hero standing tall with arms on hips, chest emblem visible, and cape flowing behind, occupying most of the vertical space.\n\nTo the right of the sub image, a compact text block labeled **\"ICONIC SILHOUETTE\"** briefly notes:  \n- **Bold Emblem:** Clear chest symbol signaling identity and purpose.  \n- **Armored Suit:** Muscular, protective design emphasizing strength.  \n- **Cape & Cowl:** Long cape and pointed cowl framing a powerful stance.\n\nBelow the sub image, a small horizontal band titled **\"CORE VISUAL ELEMENTS\"** lists three concise tags:  \n1. **STANCE:** Confident, grounded posture.  \n2. **COLOR PALETTE:** Grays and blues with a golden utility belt accent.  \n3. **FOCUS:** Minimal text, character-centered composition.\n\nIn the bottom-right corner, a subtle label reads **\"SUB IMAGE – CENTRAL HERO FIGURE\"**, reinforcing that the sub image is the main visual anchor of the slide.", "input_images": ["images/multi_image/ea664fbf-ab62-40f2-ad8e-298ef344d873_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 264, "id": "8f97790a-302a-459c-9fec-b6d78484fd14", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This infographic slide, set against a clean white background accented with whimsical blue bubbles and sparkles, utilizes a bright and engaging layout centered around the \"Despicable Me\" Minions to explain the qualities of fluffy mashed potatoes. At the very top, the title \"PERFECT MASHED POTATOES\" is displayed in large, bold, yellow bubble letters with a thick blue outline, flanked by two smiling Minions waving. Directly beneath the title is the subtitle \"SMOOTH, BUTTERY & COMFORTING!\" written in a smaller, bold, dark blue sans-serif font. The main content is organized into four distinct white rectangular panels with rounded corners and blue borders, arranged in a two-by-two grid.\n\nThe top-left panel features an image of a Minion holding a spoonful of cream (sub-image 2) on the left side. To its right, the bold heading \"WHAT ARE MASHED POTATOES?\" introduces a list of descriptive bullet points in standard black text: \"Creamy Potato Side Dish,\" \"Light, Fluffy Texture,\" \"Blended with Butter & Cream,\" and \"Seasoned with Salt & Pepper.\" The top-right panel is dedicated to \"WHY THEY’RE AWESOME\" in bold dark blue text, accompanied by an illustration of two excited Minions (sub-image 3) and floating vegetables. Below the heading, the text lists \"Pairs with Many Main Dishes,\" \"Comfort Food Favorite,\" \"Great Way to Enjoy Veggies,\" and \"Kid-Approved and Fun,\" with a smaller Minion flexing its muscles at the bottom right of the panel.\n\nMoving to the bottom row, the left panel showcases a Minion holding tacos and a baked potato (sub-image 1) next to the bold heading \"DELICIOUS WAYS TO SERVE.\" The accompanying text details its applications: \"Scooped Beside Roasts & Stews,\" \"Topped with Gravy or Cheese,\" \"Swirled with Garlic or Herbs,\" and \"Layered in Bowls & Casseroles.\" Finally, the bottom-right panel illustrates \"HOW TO MAKE THEM\" with a Minion wearing a chef's hat stirring a pot and another Minion (sub-image 4) sitting on a pile of potatoes holding a bowl of sour cream. The text in this section explains the steps: \"Peel & Boil Potatoes Until Tender,\" \"Drain and Mash While Hot,\" \"Mix In Butter, Cream & Seasonings,\" and \"Whip Until Smooth and Fluffy.\" The overall color palette is dominated by the Minion yellow, denim blue, and white, creating a cheerful and informative visual experience.", "input_images": ["images/multi_image/8f97790a-302a-459c-9fec-b6d78484fd14_input_0.jpg", "images/multi_image/8f97790a-302a-459c-9fec-b6d78484fd14_input_1.jpg", "images/multi_image/8f97790a-302a-459c-9fec-b6d78484fd14_input_2.jpg", "images/multi_image/8f97790a-302a-459c-9fec-b6d78484fd14_input_3.jpg"], "target_image": "images/multi_image/8f97790a-302a-459c-9fec-b6d78484fd14_target.png", "width": 1376, "height": 768}
+{"index": 265, "id": "0e4297ad-090d-4099-95a7-cac5846f0027", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### ROLE\nYou are a senior presentation art director and precision layout renderer.\n\n### TASK\nRender the given Structured Content JSON as one polished 16:9 presentation slide. Treat the JSON as the complete design brief: `global_style` controls the visual system, `visual_hierarchy` guides composition, and `blocks` provide the exact slide elements.\n\n### HARD RULES\n1. Use only text found in block `content`. Do not invent titles, labels, captions, summaries, or decorative copy.\n2. Follow `global_style` strictly for background, palette, typography, lighting, spacing language, and overall material feel.\n3. If `special_instructions` appears, it is the creative override. Apply it faithfully while still preserving the provided textual content.\n4. Avoid generic slides. Choose a composition that matches the content: editorial grid, hero visual, split layout, card system, timeline, layered diagram, or spacious minimalist treatment.\n5. Embed all source images into the slide.\n\n### QUALITY BAR\nThe result should look like a professionally designed keynote slide: crisp text, intentional hierarchy, balanced whitespace, coherent color usage, and no clutter.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Cozy Collaboration Dashboard\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#FDF7E9\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #FFF5D7 0%, #FDF7E9 40%, #F5E6D5 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": \\\"very subtle oversized speech bubbles in low-opacity cream behind content\\\"\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Baloo 2\\\",\\n      \\\"secondary\\\": \\\"Nunito\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#47A6B8\\\",\\n      \\\"secondary\\\": \\\"#F2C94C\\\",\\n      \\\"text\\\": \\\"#3E3A39\\\",\\n      \\\"accent\\\": \\\"#F2994A\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall look is warm and playful, echoing the soft Animal Crossing aesthetic of the central illustration. Use rounded shapes, gentle drop shadows, and speech-bubble motifs. Corners of containers should be highly rounded. The atmosphere should feel like a friendly digital workspace: bright but not harsh, with pastel yellows, teals, and warm browns that match the characters. Avoid clutter; leave generous breathing room around the hero image.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Playful, Friendly, Informal\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 28,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-28\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Friendly Feedback Hub\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#3E3A39\\\",\\n            \\\"font_family\\\": \\\"Baloo 2\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with comfortable margin above main image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\\"shadow\\\"],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"bubble\\\",\\n            \\\"special_instructions\\\": \\\"Title sits inside a pill-shaped white bubble with soft drop shadow and 24px rounded corners; bubble width about 55% of slide, centered.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/0e4297ad-090d-4099-95a7-cac5846f0027_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/0e4297ad-090d-4099-95a7-cac5846f0027_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Cute Animal Crossing-style characters collaborating at a computer with question and chat icons around them.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning 70% of slide width and anchored slightly above vertical center\\\",\\n            \\\"image_style\\\": \\\"flat illustration with no additional border\\\",\\n            \\\"container\\\": \\\"rounded panel\\\",\\n            \\\"special_instructions\\\": \\\"Place the image inside a large rounded rectangle panel with white background, 28px corner radius, and soft teal-tinted drop shadow (#47A6B8 at 15% opacity). Panel should occupy the main central area and feel like the dominant hero card.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-28-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Collect questions, chat through ideas, and mark decisions together.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Nunito\\\",\\n            \\\"color\\\": \\\"#5B5551\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image panel\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Single-line text, center-aligned, medium weight, leaving generous padding from slide edges.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-28-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A large, soft teal speech-bubble outline behind the central image, slightly offset to the right, suggesting conversation.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"decorative background shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette with #47A6B8 at low opacity\\\",\\n            \\\"position\\\": \\\"center, behind the image panel, extending beyond its edges\\\",\\n            \\\"special_instructions\\\": \\\"Opacity around 12%, thick rounded outline only (no fill), scaled so it peeks out from the top-right and bottom-left of the image container, reinforcing the communication theme without overpowering the main illustration.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate a single widescreen slide from the JSON. The output should be refined, readable, and consistent with the specified visual system.", "input_images": ["images/multi_image/0e4297ad-090d-4099-95a7-cac5846f0027_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 266, "id": "5f369271-cd47-45a8-a0b6-b74bd0a03c0c", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Set against a soft, gradient background of pale sky blue populated with stylized, fluffy white clouds, a presentation slide titled \"JOURNEY THROUGH WETLAND WORLDS\" is organized into a structured 2x2 grid. At the top center, the main title is rendered in bold, uppercase, dark blue sans-serif font, enclosed within a rounded rectangular border featuring a white fill and a blue outline. Below this header, the slide is divided into four distinct informational cards, each with a white background and rounded corners, arranged in two rows and two columns.\n\nIn the top-left quadrant, a section labeled \"LAKE LANDSCAPE\" features a thumbnail on its left side—this corresponds to the sub-image depicting a stylized 3D map of a blue lake surrounded by green hills and a small boat. To the right of this visual, the bold, black heading \"LAKE LANDSCAPE\" sits above three lines of body text in a standard black sans-serif font, reading: \"Broad, calm waters bordered by gentle slopes. Ideal for boating, fishing, and quiet retreats. A centerpiece for surrounding communities.\"\n\nDirectly to the right, in the top-right quadrant, is a section titled \"RIVER LIFELINE.\" This card displays a thumbnail on its left—matching the sub-image showing a winding blue river with fish and a rice paddy patch. To the right of the image, the bold, black heading \"RIVER LIFELINE\" is followed by two lines of text: \"Channels nutrient-rich water to farms and fields. Connects villages, markets, and distant towns.\"\n\nMoving to the bottom row, the bottom-left section is labeled \"WETLAND HABITATS.\" The visual on the left corresponds to the sub-image featuring birds flying over water with fish and lotus plants. Adjacent to this, the bold, black heading \"WETLAND HABITATS\" is placed above three lines of descriptive text: \"Shelters fish, frogs, and countless invertebrates. Rest stop for long-distance migratory birds. Marsh plants help filter and refresh the water.\"\n\nFinally, in the bottom-right corner, the section is titled \"TRADITIONS & STORIES.\" The accompanying thumbnail on the left depicts a traditional scroll with Chinese calligraphy and a dragon boat, matching the specific sub-image provided. To the right of this icon, the bold, black heading \"TRADITIONS & STORIES\" is followed by three lines of text: \"Legends of heroes and river spirits are passed down. Colorful boat races bring neighbors together in celebration.\" The entire composition utilizes a clean, educational aesthetic with a consistent color palette of blues, greens, and earth tones, emphasizing clarity and visual appeal.", "input_images": ["images/multi_image/5f369271-cd47-45a8-a0b6-b74bd0a03c0c_input_0.jpg", "images/multi_image/5f369271-cd47-45a8-a0b6-b74bd0a03c0c_input_1.jpg", "images/multi_image/5f369271-cd47-45a8-a0b6-b74bd0a03c0c_input_2.jpg", "images/multi_image/5f369271-cd47-45a8-a0b6-b74bd0a03c0c_input_3.jpg"], "target_image": "images/multi_image/5f369271-cd47-45a8-a0b6-b74bd0a03c0c_target.png", "width": 1376, "height": 768}
+{"index": 267, "id": "e7eaa49f-ad8f-434f-af23-46a84f5fab1c", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM\nYou are a cinematic presentation renderer with disciplined, editorial-grade visual design.\n\n### OBJECTIVE\nProduce a single, fully composed slide image from the input JSON while maintaining semantic precision and consistent deck-wide styling.\n\n### RENDERING POLICY\n- Text fidelity is inviolable: only display text appearing in JSON `content` fields.\n- `global_style` defines the holistic visual language, including background treatment, color system, typography, depth, lighting, and spacing.\n- `special_instructions` may refine art direction through notes on atmosphere, materials, motion or camera feel, and decorative details.\n- Preserve a clear hierarchy: titles dominate, body text remains easily readable, captions are understated, and decorative visuals must never compete with critical text.\n- Use every `source_images` entry as provided when embedding imagery in the slide.\n\n### DESIGN DIRECTION\nSelect a layout that feels distinctive yet consistent with the supplied style: strong hero focus, high-end editorial spread, sleek data infographic, modular tiles, bold centered message, or asymmetric grid. The final slide must appear sharp, legible, well-balanced, and free from any invented or omitted textual content.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Warm Urban Heat Analysis\",\n    \"background\": {\n      \"color\": \"#FFC107\",\n      \"gradient\": \"linear-gradient(180deg, #FFD54F 0%, #FFC107 100%)\",\n      \"texture\": \"noise\",\n      \"pattern\": \"subtle speckled texture overlay\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#FF5722\",\n      \"secondary\": \"#FFC107\",\n      \"text\": \"#000000\",\n      \"accent\": \"#FFD54F\"\n    },\n    \"special_instructions\": \"The slide uses a radiant, heat-inspired palette dominated by oranges and yellows. Background carries a fine grainy or speckled overlay. Large soft-edged, semi-transparent circular forms in lighter yellow/orange hues sit in opposite corners to create depth. The overall feel is urgent but analytical, ideal for highlighting environmental metrics.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Analytical, Pressing, Warm\",\n    \"composition_style\": \"Centered\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 8,\n      \"blocks\": [\n        {\n          \"id\": \"title-8\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"Tracking Bengaluru\\u2019s Rising Land Surface Temperatures\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"shadow\"\n              ],\n              \"outline_color\": null,\n              \"shadow_type\": \"drop\"\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Place the title inside a bold rounded rectangular banner with solid orange fill (#FF5722) and a pronounced black border. Center the text and scale it large for strong emphasis.\"\n          }\n        },\n        {\n          \"id\": \"illustration-8-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Large semi-transparent circular shapes in light yellow/orange tones positioned in the background corners (left and right) to add depth and visual interest.\",\n          \"style\": {\n            \"illustration_style\": \"decorative background shape\",\n            \"color_scheme\": \"light yellow and orange gradients\",\n            \"position\": \"background corners\",\n            \"special_instructions\": \"These shapes are subtle and sit behind the main content, contributing to the warm, sunny aesthetic.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_chart_box_393_212_985_478.jpg\",\n                \"relative_path\": \"imgs/img_in_chart_box_393_212_985_478.jpg\"\n              }\n            ],\n            \"source\": \"Line chart showing Max, Mean, and Min LST trends from 2016 to 2025 with peak highlights\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center, below title\",\n            \"image_style\": \"contained within a rounded rectangular frame with a light cream background\",\n            \"container\": \"rounded panel\",\n            \"special_instructions\": \"Embed the chart inside a soft-cornered white/cream panel with a gentle drop shadow so it visually lifts from the textured background.\"\n          }\n        },\n        {\n          \"id\": \"body-8-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Peak land surface temperatures in Bengaluru intensify around 2019 and 2024, reflecting rapid build-up and shrinking vegetated zones.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"below chart, centered\",\n            \"container\": \"rounded panel\",\n            \"container_opacity\": \"0.9\",\n            \"special_instructions\": \"Place this text in a light cream rounded rectangle aligned with the chart width, mirroring the chart container styling.\"\n          }\n        },\n        {\n          \"id\": \"body-8-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Shifts in minimum, mean, and maximum LST reveal how urban growth, regional climate trends, and cooling initiatives interact over time.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom, centered\",\n            \"container\": \"rounded panel\",\n            \"container_opacity\": \"0.9\",\n            \"special_instructions\": \"Set this secondary message in a matching cream rounded box positioned beneath the first body block, maintaining consistent margins and alignment.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nCreate the final 16:9 slide image from the JSON above, following all style and content rules exactly.", "input_images": ["images/multi_image/e7eaa49f-ad8f-434f-af23-46a84f5fab1c_input_0.jpg"], "target_image": "images/multi_image/e7eaa49f-ad8f-434f-af23-46a84f5fab1c_target.png", "width": 1376, "height": 768}
+{"index": 268, "id": "0d7a5756-8fd5-4954-adf3-2785fd840962", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "Render one presentation slide from the provided Structured Content JSON.\n\nThe JSON is the single source of truth:\n1. Use `global_style` for the slide's visual system.\n2. Use `visual_hierarchy` for emphasis, mood, and composition.\n3. Use `blocks` for exact text and visual elements.\n4. Use `special_instructions` as a priority creative override.\n5. Use `source_images` to embed in the slide.\n\nNever add or rewrite textual content. You may only improve typography through layout decisions such as line breaks, alignment, font weight, spacing, and scale.\n\nMake the slide feel designed, not templated. Avoid monotonous bullet layouts. Use thoughtful whitespace, clean grouping, precise alignment, and a composition that matches the content.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Quiet Woodland Liturgy\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#101713\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top, #253328 0%, #0B1110 65%, #050708 100%)\\\",\\n      \\\"texture\\\": \\\"paper\\\",\\n      \\\"pattern\\\": \\\"Very subtle mottled texture evoking aged parchment, barely visible behind the content.\\\"\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Cinzel\\\",\\n      \\\"secondary\\\": \\\"Cormorant Garamond\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#D9C9A1\\\",\\n      \\\"secondary\\\": \\\"#4A6B52\\\",\\n      \\\"text\\\": \\\"#F4F1E9\\\",\\n      \\\"accent\\\": \\\"#B78B4B\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall mood is contemplative and reverent, inspired by an outdoor religious ritual. The palette blends deep forest greens and warm parchment tones to echo the robe and the surrounding trees. Lighting should feel soft and diffused, as if filtered through leaves, with gentle vignetting around the slide edges to draw attention toward the center. Subtle gold accents recall illuminated manuscripts without overpowering the understated atmosphere. Decorations, if any, remain minimal: thin lines and small crosses or leaf motifs, rendered delicately.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Serene, Reverent, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 30,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/0d7a5756-8fd5-4954-adf3-2785fd840962_input_0.png\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/0d7a5756-8fd5-4954-adf3-2785fd840962_input_0.png\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"A person in white liturgical robes holding a book in a green forest setting.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, occupying roughly 70% of slide height with generous margins on left and right\\\",\\n            \\\"image_style\\\": \\\"Soft drop shadow with a thin, off-white border\\\",\\n            \\\"container\\\": \\\"Subtle rounded-rectangle frame in very dark green (#0B1110) slightly lighter than the background\\\",\\n            \\\"special_instructions\\\": \\\"Apply a slight vignette around the photo edges to blend into the dark frame; use a soft 12px border-radius. Ensure the image remains the clear focal point, with the frame width about 65% of the slide width.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"title-30\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Moments of Quiet Devotion\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#D9C9A1\\\",\\n            \\\"font_family\\\": \\\"Cinzel\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, above the image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Set in small caps with slightly increased letter spacing. Align centered, leaving ample negative space between the title and the top edge as well as between the title and the image.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-30-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"A quiet ceremony unfolds beneath the shelter of the trees.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Cormorant Garamond\\\",\\n            \\\"color\\\": \\\"#F4F1E9\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Set in italic, medium size, centered. Keep the line short and allow additional breathing room to the slide bottom edge.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-30-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A pair of very thin horizontal lines with a small golden cross-like motif at the center, placed beneath the title as a decorative divider.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"line art\\\",\\n            \\\"color_scheme\\\": \\\"Uses accent gold (#B78B4B) and soft parchment (#D9C9A1)\\\",\\n            \\\"position\\\": \\\"top-center, directly between the title and the top edge of the image\\\",\\n            \\\"special_instructions\\\": \\\"Lines should be subtle, approximately 40% of slide width, with the tiny cross or floral symbol at their midpoint. Low opacity (~0.6) so they remain understated and do not compete with the main image.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nRender this JSON as a single 16:9 slide image. Make it visually compelling, readable, and consistent with all provided styling constraints.", "input_images": ["images/multi_image/0d7a5756-8fd5-4954-adf3-2785fd840962_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 269, "id": "b8dcdb6c-f338-4d4a-8c2d-9f128624a4d4", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, titled **\"PROGRAM SEGMENT FREQUENCY PATTERNS,\"** uses observed data to highlight how often different program segments appear under varying conditions. The slide centers on a scatter plot illustrating the relationship between frequency and the number of program segments, with distinct markers for two categories of behavior. The first sub-image, the scatter plot with blue and orange points, corresponds to the central visual element of the slide and is labeled **\"SEGMENT FREQUENCY VS. PROGRAM COUNT (OBSERVED DATA).\"** Blue dots denote **Non-attack** segments, while orange dots indicate **Attack** segments, as shown in the legend on the upper right of the sub image. The horizontal axis represents **Frequency**, and the vertical axis represents **# Program segments**, emphasizing how attacks cluster differently from non-attacks.\n\n**KEY OBSERVATIONS**\n- Non-attack segments show a wide spread of frequencies and counts.  \n- Attack segments cluster at lower frequencies and program counts.  \n\nA footer at the bottom reads **\"PATTERNS ARE ILLUSTRATIVE – FOR ANOMALY INSIGHT ONLY.\"**", "input_images": ["images/multi_image/b8dcdb6c-f338-4d4a-8c2d-9f128624a4d4_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 270, "id": "9383239f-2105-4c82-b31a-521d3253fa03", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### MISSION\nTransform the input Structured Content JSON into one elegant, high-fidelity presentation slide.\n\n### STRICT CONTENT CONTROL\nOnly the JSON may supply visible content. Do not add explanatory copy, generated bullets, fake labels, placeholder text, or decorative words. If text is sparse, preserve the sparse feeling.\n\n### STYLE CONTROL\nHonor `global_style` as the governing design system. Use the specified background, color palette, typography, and overall visual mood. Apply `special_instructions` when provided, especially for atmosphere, lighting, texture, or layout emphasis.\n\n### BLOCK HANDLING\nRender visible blocks according to their role and hierarchy. Use `script` only as hidden context for the slide's theme and emotional tone. Keep text blocks readable and avoid cramped paragraphs.\n\n### IMAGE INTEGRITY & PRESERVATION\nWhen you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n\n### LAYOUT EXPECTATION\nDesign a slide with deliberate composition. Vary structure across slides: some may be visual-led, some text-led, some diagrammatic, some minimal. The output should feel premium, consistent, and ready for a professional deck.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Deep Learning Training Dynamics\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#050910\\\",\\n      \\\"gradient\\\": \\\"radial gradient from #0B1220 at center to #020308 at edges\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": \\\"subtle diagonal grid lines in very low-opacity #1E293B to evoke a technical dashboard\\\"\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"IBM Plex Sans\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#38BDF8\\\",\\n      \\\"secondary\\\": \\\"#FBBF24\\\",\\n      \\\"text\\\": \\\"#E5E7EB\\\",\\n      \\\"accent\\\": \\\"#22C55E\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall atmosphere is modern, technical, and focused. Dark data-viz style background with neon-like blue and amber accents that echo the lines in the chart. Clean sans-serif typography, thin dividing lines, and subtle glow around key elements to highlight the analytical feel. No clutter; plenty of breathing room around the central figure.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Analytical, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 16,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-16\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Stability of DLLM vs AR Training\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#E5E7EB\\\",\\n            \\\"font_family\\\": \\\"IBM Plex Sans\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center with generous margin above the chart\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Title in all caps or small caps, tracking slightly expanded. A soft cyan glow behind the text to tie into the chart colors.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/9383239f-2105-4c82-b31a-521d3253fa03_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/9383239f-2105-4c82-b31a-521d3253fa03_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Training loss and gradient norm curves for DLLM and AR models.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning about 75% of slide width and 65% of height\\\",\\n            \\\"image_style\\\": \\\"subtle outer glow with thin light border\\\",\\n            \\\"container\\\": \\\"rounded rectangle frame\\\",\\n            \\\"special_instructions\\\": \\\"Use a 2px border in #1F2937 with 16px corner radius. Apply a faint cyan outer glow on the frame to make the image pop against the dark background. The figure should be the dominant element, horizontally centered with equal left/right margins and slightly more space below than above.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-16-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"DLLM shows higher variance in gradient norms, while AR converges with smoother, smaller updates.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#9CA3AF\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, directly beneath the image frame\\\",\\n            \\\"container\\\": null,\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Text left-aligned within a narrow width centered under the chart. Use small font size with slightly increased line spacing for readability.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-16-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two thin, semi-transparent gradient arcs curving upward from the bottom corners toward the sides of the chart, suggesting data flow.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette with cyan and amber fading into transparent\\\",\\n            \\\"position\\\": \\\"background layer near the lower left and right edges of the slide\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (around 20%), placed behind all text and image. Arcs should be soft and subtle, not distracting, reinforcing the technical and data-driven mood.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nProduce a polished widescreen slide image from the JSON above, balancing style consistency with a visually distinct composition.", "input_images": ["images/multi_image/9383239f-2105-4c82-b31a-521d3253fa03_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 271, "id": "6320eb81-0f6a-4383-acb0-67f9613a3d08", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The slide illustrates a clean, technical infographic titled **\"DIFFUSION CAUSAL INTERVENTION\"** displayed in large, bold, black sans‑serif capital letters at the very top center. The background is a light, slightly textured grey with a subtle grid pattern, giving a modern, research‑poster feel. Thin, charcoal‑colored borders frame the slide, with small, minimal icons (simple circles and squares) decorating the four corners to echo a schematic, diagrammatic style.\n\nCentered prominently in the upper half of the slide is the **sub image** itself, occupying most of the width. It is shown without cropping, surrounded by a thin, dark grey outline to make it stand out as the main focus. Directly beneath this frame, in small uppercase text, is a short label: **\"SUB IMAGE: CAUSAL INTERVENTION OVER DIFFUSION STEPS\"**.\n\nThe lower half of the slide is divided into **three compact text regions**, each using minimal wording to support the visual:\n\n1. **Left block – \"SETUP\"**  \n   - Heading: **\"1. SETUP\"** in bold black small caps, aligned left.  \n   - Beneath the heading are two ultra‑short bullet points, in a smaller sans‑serif font:\n     - \"Original diffusion pipeline with prompt tokens.\"\n     - \"Clean vs. modified encoder states (E\\_CLEAN, E\\_MODIFIED).\"  \n   - A thin vertical rule on the right separates this block from the central area, matching the schematic lines seen in the sub image.\n\n2. **Center block – \"INTERVENTION\"**  \n   - Heading: **\"2. INTERVENTION ON PADS\"** centered, bold.  \n   - Under the heading is a very brief two‑line description, centered:\n     - \"Swap contextual pad tokens during diffusion steps.\"\n     - \"Trace causal impact on the generated scene.\"  \n   - Directly below, a small horizontal arrow icon points upward toward the middle portion of the **sub image**, visually reinforcing that this text describes the blue “Causal Intervention Diffusion” band and the pad tokens shown there.\n\n3. **Right block – \"RESULT\"**  \n   - Heading: **\"3. VISUAL EFFECT\"** in bold, aligned right.  \n   - Two short bullets, right‑aligned:\n     - \"Left: image from intervened pads.\"\n     - \"Right: original reconstruction.\"  \n   - A short note in tiny text beneath, still right‑aligned:  \n     - \"Compare how prompt‑context pads steer the compass imagery.\"\n\nAcross the very bottom of the slide, centered in a subtle, light‑grey small‑caps font, is a closing line: **\"FOCUS: HOW PAD TOKEN INTERVENTIONS CHANGE DIFFUSION OUTPUTS\"**. All textual elements remain secondary in size and weight to the **sub image**, ensuring that the viewer’s attention is drawn first and foremost to the diagram of the diffusion causal intervention process.", "input_images": ["images/multi_image/6320eb81-0f6a-4383-acb0-67f9613a3d08_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 272, "id": "88562c68-d4fe-4d89-bd61-42eb72dd8c5e", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "With a clean, structured layout, this presentation slide titled **\"RIVER TRANQUILITY\"** in bold, dark text at the top highlights the calming essence of flowing water and lush greenery. The slide is divided into four vertical sections, each headed by a simple circular icon. The provided **sub image** of the forest stream is placed prominently in the center, spanning the upper half of the slide to serve as the visual focal point.\n\nThe first column on the left is labeled **\"SCENE\"** and lists short descriptors:  \n- \"FOREST STREAM\"  \n- \"MOSSY BOULDERS\"  \n- \"CLEAR, RUSHING WATER\"  \n\nThe second column, titled **\"MOOD\"**, contains concise phrases:  \n- \"CALMING FLOW\"  \n- \"GENTLE ENERGY\"  \n- \"NATURAL HARMONY\"  \n\nThe third column, named **\"DETAILS\"**, briefly notes:  \n- \"TEXTURED ROCKS\"  \n- \"RICH GREEN FOLIAGE\"  \n- \"LAYERED CASCADES\"  \n\nThe fourth column on the right is labeled **\"USE & INSPIRATION\"** and includes:  \n- \"WELLNESS & MINDFULNESS\"  \n- \"NATURE THEMES\"  \n- \"BACKGROUND VISUALS\"  \n\nA cool palette of greens and soft blues echoes the tones of the **sub image**, with light geometric accents to maintain a modern, uncluttered feel. A subtle footer in the bottom right reads **\"NATURE SCENES | STREAM SERIES\"**.", "input_images": ["images/multi_image/88562c68-d4fe-4d89-bd61-42eb72dd8c5e_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 273, "id": "1c11d67e-632e-446a-a7e5-61b7e5062f76", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & PURPOSE\nYou are an elite AI Slide Visualization Engine, fusing the sensibilities of a Senior Medical Illustrator with the rigor of an Information Designer.\n\n**Primary Mission:** Transform the incoming **Structured Content JSON** into a single, production-ready 16:9 presentation slide image.\n\n### INPUT DATA SEMANTICS\nA single-slide JSON definition will be provided. Interpret each key as:\n\n1. **`global_style`** – The non-negotiable global art direction: background treatment, typography system, and color language.\n2. **`visual_hierarchy`** – The compositional rhythm and intended emphasis.\n3. **`blocks`** – The discrete renderable objects on the slide (text fields, diagrams, embedded images, decorative elements).\n\n### EXECUTION STANDARDS (MANDATORY)\n\n#### 1. EMBEDDED IMAGE FIDELITY (TOP PRIORITY)\n* **Directive:** Any block with `type: \"image\"` must render the exact supplied image file(s) in the final slide.\n* **Image Assets:** Real image files are provided as attachments. These must be used as-is.\n* **Preservation Rules:**\n  * Do **NOT** crop, scale non‑uniformly, recolor, or conceptually alter the pixel content of the images.\n  * Do **NOT** substitute with AI-generated art, icons, or placeholder stock visuals.\n  * Do **NOT** omit or hide any of the provided image assets.\n* **Permitted Treatments:** You may stylize the *surrounding frame* (e.g., soft borders, subtle drop shadows, reflections, perspective placement) provided the underlying image remains perfectly readable and undistorted.\n* **Multiple Image Handling:** When several images are present, lay them out as indicated in the JSON or, if unspecified, choose a balanced composition (row, column, or grid) consistent with `visual_hierarchy`.\n\n#### 2. TEXT SOURCING & LIMITS\n* **Text Source of Truth:** Only render text that appears in the JSON `content` fields of text-based blocks.\n* **Restrictions:**\n  * **No Fabrication:** Do not create extra titles, paragraphs, labels, or annotations that are not explicitly contained in the JSON.\n  * **No Padding Copy:** Resist the urge to fill empty zones with decorative or explanatory text.\n* **Typographic Control:** You may reflow and style text (line breaks, font size within its role, emphasis, spacing) for aesthetics and legibility, but avoid dense, unbroken paragraphs.\n* **`role: \"script\"` Blocks:** Treat these strictly as background context for mood and style decisions. Do **not** display script text on the slide unless `special_instructions` explicitly instruct you to.\n\n#### 3. SPECIAL_INSTRUCTIONS AS CREATIVE MASTER SWITCH\n* **Override Hierarchy:** Whenever a `special_instructions` field exists (in `global_style` or within an individual block), its directives supersede general rules, unless they conflict with image fidelity or text-source rules.\n  * Example: If the base style calls for a flat background but `special_instructions` requests “soft volumetric lighting with subtle glow,” prioritize the latter.\n* **Usage:** Employ `special_instructions` to refine textures, lighting, micro‑layout, motion cues (visually implied), or to emphasize particular elements.\n\n#### 4. GLOBAL_STYLE ENFORCEMENT (CROSS-SLIDE COHERENCE)\n* **Non-Variable Elements:**\n  * **Background:** Use the exact `background.color`, gradients, and described pattern. This must remain visually identical wherever this style is applied.\n  * **Typography:** Titles use `font_palette.primary`; body and supporting text use `font_palette.secondary`.\n  * **Color Mapping:** Whenever `color_ref` or palette keys are referenced, apply the corresponding hex codes accurately.\n* **Role-Based Typographic Consistency:**\n  * All title elements with `importance: \"dominant\"` share the same approximate point size, weight, and canonical position across slides.\n  * All supporting or body text with subheading/standard roles must be consistent in font size and styling.\n  * Caption or note text must keep a uniform, smaller caption style.\n* **Intent:** Enable independently rendered slides to assemble into a visually unified deck.\n\n#### 5. COMPOSITION VARIETY & INFORMATION DESIGN (IMPORTANT)\n* **Avoid a Single Template:** Do not default to repetitive “header on top + bullets below” unless explicitly requested.\n* **Layout Strategies:**\n  * **Dense text:** Use multi-column layouts, staggered boxes, or asymmetry rather than a single vertical stack.\n  * **Graphic-led:** When illustrations or images dominate, allow them to command the central space. Position text as concise, supporting labels or callouts.\n  * **Hybrid:** Balance visual elements and text with overlapping panels, grouped containers, or zoned regions.\n  * **Whitespace:** Where content is minimal, preserve generous negative space, enhancing focus and perceived quality rather than inflating elements unnaturally.\n* **Variation Tools:**\n  * Shift text alignment (left, right, center) and anchoring across slides.\n  * Use different panel shapes (rounded cards, pill banners, floating chips) while keeping colors and fonts consistent.\n  * Adjust weight distribution so some slides are title‑centric while others are illustration‑centric.\n\n### RENDERING OUTPUT\nProduce a single 16:9 slide image that visually implements the provided JSON and adheres to all above policies.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Medical Educational Infographic\",\n    \"background\": {\n      \"color\": \"#2B72C3\",\n      \"gradient\": \"linear-gradient(135deg, #2B72C3, #1A5BA8)\",\n      \"texture\": \"none\",\n      \"pattern\": \"Abstract soft curved shapes in lighter blue tones (#4A90E2) layered in the background\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#F4C430\",\n      \"secondary\": \"#2B72C3\",\n      \"text\": \"#FFFFFF\",\n      \"accent\": \"#FF6B6B\"\n    },\n    \"special_instructions\": \"The slide features a clean, modern medical infographic style. The background is a rich blue with subtle abstract wave patterns. Key visual elements use rounded corners and soft drop shadows to create depth. The color scheme is high-contrast with bright yellow and white text against the blue background for readability. Medical illustrations are rendered in a realistic but stylized manner with soft shading.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Educational, Clinical\",\n    \"composition_style\": \"Centered, Hierarchical with Top-Down Flow\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 11,\n      \"blocks\": [\n        {\n          \"id\": \"title-11\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"INSIDE THE BRAIN: PARENCHYMAL INJURY PATTERNS\",\n          \"style\": {\n            \"color\": \"#2B72C3\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Text is contained within a bright yellow (#F4C430) rounded rectangular banner that spans most of the slide width. The banner has a subtle drop shadow.\"\n          }\n        },\n        {\n          \"id\": \"illustration-11-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A composite medical diagram showing three views of the brain: a sagittal cross-section in the center showing internal structures, a top-down view of the brain on the right, and a magnified inset on the left showing microbleeds. The central brain diagram has red shaded areas indicating injury sites in the frontal lobe, temporal lobe, and brainstem. Arrows indicate 'Coup Mechanism' and 'Contrecoup Mechanism'. A circular magnified view shows the corpus callosum with purple dots indicating DAI (Diffuse Axonal Injury).\",\n          \"style\": {\n            \"illustration_style\": \"medical diagram with labels and callouts\",\n            \"color_scheme\": \"pinkish-beige brain tissue, red injury highlights, purple dots for DAI, black text labels\",\n            \"position\": \"center, spanning horizontally across the middle of the slide\",\n            \"special_instructions\": \"The illustration includes multiple callout lines with labels pointing to specific brain regions: Frontal Lobe, Temporal Lobe, Brainstem, Corpus Callosum. There are curved arrows indicating force mechanisms. The style is semi-realistic with soft shading.\"\n          }\n        },\n        {\n          \"id\": \"image-11-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_148_208_339_411.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_148_208_339_411.jpg\"\n              }\n            ],\n            \"source\": \"Magnified inset showing microbleeds in brain tissue\",\n            \"caption\": \"MICROBLEEDS (high T2* = contusion)\"\n          },\n          \"style\": {\n            \"position\": \"left side, vertically centered, connected to main brain diagram with a light blue cone shape\",\n            \"image_style\": \"contained within a white rectangular frame with rounded corners\",\n            \"container\": \"white rounded rectangle with subtle shadow\",\n            \"special_instructions\": \"This inset image is a zoomed-in view of the brain tissue showing red spots representing microbleeds. It is connected to the main brain diagram by a light blue conical shape indicating magnification.\"\n          }\n        },\n        {\n          \"id\": \"body-11-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Contusions and diffuse axonal injury are the most common brain parenchymal injuries in TBI.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left\",\n            \"container\": \"dark blue rounded rectangle with subtle shadow\",\n            \"container_opacity\": \"0.9\",\n            \"special_instructions\": \"Text is preceded by a small yellow circular bullet point. The container has a slight gradient and rounded corners.\"\n          }\n        },\n        {\n          \"id\": \"body-11-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Advanced MRI is critical for detecting subtle but severe injuries such as DAI.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right\",\n            \"container\": \"dark blue rounded rectangle with subtle shadow\",\n            \"container_opacity\": \"0.9\",\n            \"special_instructions\": \"Text is preceded by a small teal circular bullet point. The container has a slight gradient and rounded corners, matching the style of the left text box.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\n\nGenerate a polished 16:9 slide image that strictly follows this JSON configuration and incorporates every attached source image without alteration or omission.", "input_images": ["images/multi_image/1c11d67e-632e-446a-a7e5-61b7e5062f76_input_0.jpg"], "target_image": "images/multi_image/1c11d67e-632e-446a-a7e5-61b7e5062f76_target.png", "width": 1376, "height": 768}
+{"index": 274, "id": "3d2e7f33-8d57-45ec-b25d-5f5ba79e3c78", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The presentation slide is titled \"DIGITAL TRANSFORMATION FRAMEWORK\" in large, black, uppercase sans-serif font, anchored at the top left by a vertical light blue bar. The layout is divided into two main sections: a three-column framework explanation on the left, and a right-hand column featuring two data visualizations and a concise footer. The background is a clean, solid light gray-blue.\n\nThe left section introduces three strategic domains using icons and bulleted lists. The first column, \"CUSTOMER EXPERIENCE,\" features an icon of a stylized user silhouette connected to a chat bubble and a star, all supported by an open hand. Below this icon, the title \"CUSTOMER EXPERIENCE\" appears in bold black uppercase letters, followed by four bullet items in standard black text: \"Omnichannel Interactions,\" \"Personalized Journeys,\" \"Proactive Support & Service,\" and \"Feedback-Driven Improvements.\" The central column, \"OPERATIONAL EXCELLENCE,\" displays an icon group comprising interlocking gears, a speedometer, and a checklist. Beneath this, the bolded title \"OPERATIONAL EXCELLENCE\" is followed by the bullet points: \"Process Automation & Workflow,\" \"Data-Driven Decision Making,\" \"Continuous Improvement,\" and \"Risk Management & Compliance.\" The third column, \"INNOVATION & LEADERSHIP,\" shows an icon of a lightbulb, a rocket, and an organizational chart. Under the bolded header \"INNOVATION & LEADERSHIP,\" the list reads: \"Agile Culture & Mindset,\" \"Cross-Functional Collaboration,\" \"Strategic Technology Investments,\" and \"Visionary Leadership & Governance.\" The input image corresponding to the bar chart is placed on the right side of the slide.\n\nThe right side of the slide features two charts. The upper chart is a donut chart titled \"DIGITAL INVESTMENT DISTRIBUTION (Illustrative).\" The chart is segmented into three colored sections: a green section labeled \"Platform Modernization 40%,\" a blue section labeled \"Customer-Facing Solutions 35%,\" and a teal section labeled \"Data & Analytics 25%.\" The input image showing the pie chart is placed on the right side of the slide. Below this is a stacked bar chart titled \"MIX OF BUSINESS ACTIVITIES OVER TIME (Example Scenario).\" The chart displays three bars labeled \"Year 1,\" \"Year 3,\" and \"Year 5\" along the x-axis. A legend on the right identifies the stack segments by color: a green square for \"Digital-First Initiatives,\" a blue square for \"Enhanced Legacy Processes,\" and a grey square for \"Traditional Operations.\" The bars show a trend where the grey \"Traditional Operations\" segment shrinks over time while the green \"Digital-First Initiatives\" segment grows. At the very bottom right corner of the slide, the footer text reads \"CLEAR | MODERN | FUTURE-READY\" in small, light grey uppercase letters. The input image showing the corporate pillars is placed on the left side of the slide.", "input_images": ["images/multi_image/3d2e7f33-8d57-45ec-b25d-5f5ba79e3c78_input_0.jpg", "images/multi_image/3d2e7f33-8d57-45ec-b25d-5f5ba79e3c78_input_1.jpg"], "target_image": "images/multi_image/3d2e7f33-8d57-45ec-b25d-5f5ba79e3c78_target.png", "width": 1376, "height": 768}
+{"index": 275, "id": "4417b872-9693-4022-91e1-3bb8931c0c7a", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide adopts a retro, 1930s animation aesthetic, utilizing a limited color palette of terracotta red, slate blue, cream, and black against a white background. The layout is framed by a decorative border featuring geometric shapes and repeating wave motifs in the corners. In the bottom-left corner, a small text label reads \"1930s ANIMATION\" in a black, sans-serif font.\n\nAt the top left, the main title is displayed in a large, bold, terracotta-colored serif font with a black outline, reading \"SUNLIT BLOOMS:\". Directly beneath it, in a smaller, bold, black sans-serif font, is the subtitle \"A MOMENT IN YELLOW\".\n\nThe slide is visually divided into two main sections. On the left side, occupying most of the vertical space, a large rectangular frame contains the sub image: a dense cluster of bright yellow daisy-like flowers with dark centers, their petals radiating outward in overlapping layers. Thin green stems and leaves weave through the blossoms, adding contrast. The photograph is presented without additional effects so that the natural colors remain vivid. A slim slate-blue inner border outlines the photo, giving it a gallery-style appearance. Beneath this graphic, centered in bold, terracotta letters, the caption reads \"FIELD OF FLOWERS\".\n\nTo the right, the layout emphasizes minimal text and generous white space, keeping the sub image as the focal point. Near the upper-right corner, a section header in bold, terracotta serif text reads \"VISUAL NOTES\". Below this header, three short, vertically stacked text lines are set in a clean black sans-serif font, each preceded by a small geometric icon in slate blue:\n\n1. The first line reads \"COLOR\" with a brief note underneath: \"Bright yellow petals & deep centers\".\n2. The second line reads \"TEXTURE\" with the note: \"Soft layers of overlapping blooms\".\n3. The third line reads \"MOOD\" with the note: \"Warm, lively, sunlit atmosphere\".\n\nA small cream-colored panel extends from the right margin toward the center, echoing the look of an old animation title card. Within this panel, a single succinct sentence in black sans-serif type states: \"The sub image captures a close-up view of cheerful, golden flowers in full bloom.\"\n\nBeneath these notes, aligned near the bottom-right, a modest decorative motif of terracotta and slate-blue leaves and petals mirrors the shapes seen in the sub image, subtly tying the graphic elements together without drawing attention away from the photograph.\n\nThe entire composition is enclosed within a border composed of horizontal lines with geometric interruptions and corner accents matching the \"1930s ANIMATION\" style, ensuring that the sub image remains the dominant visual element while the surrounding design gently frames and enhances it.", "input_images": ["images/multi_image/4417b872-9693-4022-91e1-3bb8931c0c7a_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 276, "id": "8b0bd26f-6b11-4279-8fa2-98d32a837c86", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide, titled **\"OCEAN FORCE & COASTAL ROCKS\"** in large, bold, black, uppercase letters centered at the top, is designed with a clean white background and organized into four distinct quadrants. Thin, dashed dark‑blue vertical and horizontal lines divide the slide into a structured 2x2 grid. The visual palette is minimal, using black for text, deep blue for dividers and icons, and subtle accents of grey and sea‑green.\n\nIn the **top-left quadrant**, under the bold heading **\"SCENE OVERVIEW\"**, a small dark‑blue circular icon of a simplified wave and rock sits to the left of the text. The text content is minimal, consisting of three short bullet points in black:  \n- “Rocky shoreline”  \n- “Breaking waves”  \n- “Open sea horizon”  \n\nIn the **top-right quadrant**, labeled **\"SUB IMAGE FOCUS\"** in bold, the main sub image is placed prominently, occupying most of the quadrant. The sub image shows a powerful wave crashing against jagged coastal rocks, sending a white spray upward while the textured ocean extends toward a bright, shimmering horizon. A thin dark‑blue border frames the sub image, making it the central visual element of the slide. Below the sub image, in small grey uppercase text, is a caption: **“SUB IMAGE: WAVE IMPACT MOMENT”**.\n\nThe **bottom-left quadrant** is titled **\"KEY VISUAL ELEMENTS\"** in bold. To the left of the text is a modest icon: three layered lines suggesting ocean waves. The textual content, kept concise, highlights what to notice in the sub image, using numbered points in black:  \n1. **TEXTURED WATER SURFACE** – ripples and deeper blue tones.  \n2. **SPRAYING WAVE PLUME** – bright white, mid‑crash.  \n3. **DARK COASTAL ROCKS** – rough shapes framing the foreground.\n\nFinally, the **bottom-right quadrant** bears the heading **\"MOOD & USE\"** in bold, with a circular icon beside it depicting a small sun above stylized waves. The text, arranged as three short bullets, suggests how the sub image might be used or interpreted:  \n- “Express power of nature”  \n- “Backdrop for marine topics”  \n- “Calm yet dramatic atmosphere”\n\nAll headings are in uppercase and bold, clearly aligned just above their respective content areas. Body text is left‑aligned within each quadrant, leaving ample white space so the viewer’s attention remains centered on the sub image of the crashing wave and rugged coast.", "input_images": ["images/multi_image/8b0bd26f-6b11-4279-8fa2-98d32a837c86_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 277, "id": "8d8b68c3-5ca3-4161-abac-d879527c2b00", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "Render one presentation slide from the provided Structured Content JSON.\n\nThe JSON is the single source of truth:\n1. Use `global_style` for the slide's visual system.\n2. Use `visual_hierarchy` for emphasis, mood, and composition.\n3. Use `blocks` for exact text and visual elements.\n4. Use `special_instructions` as a priority creative override.\n5. Use `source_images` to embed in the slide.\n\nNever add or rewrite textual content. You may only improve typography through layout decisions such as line breaks, alignment, font weight, spacing, and scale.\n\nMake the slide feel designed, not templated. Avoid monotonous bullet layouts. Use thoughtful whitespace, clean grouping, precise alignment, and a composition that matches the content.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Calm Workspace Focus\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#F4F7FB\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #FFFFFF 0%, #F4F7FB 40%, #E4ECF5 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Nunito\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#4F7CAC\\\",\\n      \\\"secondary\\\": \\\"#9BB6D1\\\",\\n      \\\"text\\\": \\\"#2B3A42\\\",\\n      \\\"accent\\\": \\\"#88B04B\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall aesthetic is soft, bright, and professional, evoking a calm home-office setting. Use plenty of negative space around the central image. Corners are gently rounded and shadows are subtle, as if the content card floats above the pale blue background. Accents in muted blue and soft green echo the sweater, plants, and mug from the illustration. Typography is clean and friendly with medium-rounded sans-serif fonts, maintaining readability while feeling approachable.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Calm, Focused\\\",\\n    \\\"composition_style\\\": \\\"Centered hero with title above and subtle subtitle below\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 7,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-7\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Deep Work in a Digital World\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#4F7CAC\\\",\\n            \\\"font_family\\\": \\\"Nunito\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with generous margin above image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"none\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Set in large, bold letters centered across the slide. Tracking slightly expanded for an airy feel. Leave ample white space between the title and the image below.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/8d8b68c3-5ca3-4161-abac-d879527c2b00_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/8d8b68c3-5ca3-4161-abac-d879527c2b00_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Illustration of a person working at a laptop with documents, coffee, and plants on the desk.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, slightly below midline, spanning about 70% of slide width\\\",\\n            \\\"image_style\\\": \\\"soft drop shadow with rounded corners\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Apply a 16–24px border-radius to echo the friendly style of the illustration. Use a very subtle, blurred shadow (#000000 at 10–15% opacity) to lift the image from the background. Ensure the image is the largest visual element on the slide.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-7-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Balancing focused thinking, clear notes, and a touch of green on your desk.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#2B3A42\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image within the lower third\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Set in small to medium size, centered. Use regular weight for a gentle, explanatory tone. Maintain sufficient spacing from the slide edge to keep the layout light.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-7-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Two soft, overlapping circular gradient shapes behind the lower corners of the main image, hinting at abstract desk lights or focus halos.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette with pale blue and soft green tints\\\",\\n            \\\"position\\\": \\\"one shape behind the bottom-left of the image, the other behind the bottom-right, partially off-screen\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (around 20–30%), large radius circles softly blurred. Left circle in light blue (#D7E3F4), right circle in pale green (#C9E3B0). They must remain subtle so they don’t compete with the central illustration.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nCreate the final 16:9 slide image from the JSON above, following all style and content rules exactly.", "input_images": ["images/multi_image/8d8b68c3-5ca3-4161-abac-d879527c2b00_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 278, "id": "ed23cb09-7a16-42c1-ac42-bfd3f409b187", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, titled \"DIGITAL CAMPAIGN LIFECYCLE & PLAN\" in large uppercase letters at the top, outlines a framework for managing a marketing initiative. On the left, under the heading \"GOALS OF THIS CAMPAIGN\", a list of five objectives is presented with accompanying teal icons: \"BOOST BRAND VISIBILITY\", \"DRIVE QUALIFIED LEADS\", \"ALIGN MESSAGING ACROSS CHANNELS\", \"MAXIMIZE CONTENT IMPACT\", and \"IMPROVE AUDIENCE RETENTION\". The right side features the \"CAMPAIGN EXECUTION ROADMAP,\" which consists of a horizontal timeline bar broken into four segments: \"Weeks 1-2\", \"Weeks 3-6\", \"Weeks 7-8\", and \"Ongoing\". Below this timeline, a cascading Gantt chart details four sequential steps: \"PHASE 1: STRATEGY & DISCOVERY\" (light green), \"PHASE 2: ASSET CREATION & REFINEMENT\" (medium green), \"PHASE 3: APPROVALS & LAUNCH\" (teal), and \"PHASE 4: ANALYTICS & OPTIMIZATION\" (dark teal arrow). The sub-image showing the detailed timeline and phases corresponds to the roadmap section on the right, while the full slide screenshot provides the complete layout with the \"GOALS OF THIS CAMPAIGN\" section on the left.", "input_images": ["images/multi_image/ed23cb09-7a16-42c1-ac42-bfd3f409b187_input_0.jpg"], "target_image": "images/multi_image/ed23cb09-7a16-42c1-ac42-bfd3f409b187_target.png", "width": 1376, "height": 768}
+{"index": 279, "id": "afbbaa54-c08b-4ad0-a32e-254b0ca6e80a", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Displaying a futuristic, high-tech slide titled \"COLLABORATIVE INTELLIGENCE: A CONNECTED WORKSPACE\" in bold, white, sans-serif uppercase letters located in the top-left corner. The background is a deep navy blue overlaid with a faint, circuit-board-style grid pattern in varying shades of cyan and dark blue, giving the impression of a digital interface. In the top-right corner, small line-art icons of gears and a brain float subtly. The main body of the slide features four vertical rectangular panels with glowing cyan borders, arranged horizontally to represent a process flow. From left to right, the first panel is labeled \"INSIGHT DISCOVERY\" at the top and repeats the phrase \"INSIGHT DISCOVERY\" below a central icon of a globe connected to nodes. Below this, a list reads: \"SCAN SIGNALS & TRENDS\", \"AGGREGATE RAW DATA\", \"EXTRACT MEANINGFUL PATTERNS\". The second panel, glowing with an orange gradient at the bottom, is titled \"STRUCTURING & CURATION\" and features an icon of interconnected folders. The text below repeats \"STRUCTURING & CURATION\" followed by: \"BUILD LOGICAL TAXONOMIES\", \"GROUP RELATED ITEMS\", \"TAG WITH CONTEXTUAL LABELS\". The third panel, also with an orange glow, is titled \"ARCHIVING & ACCESS\" and displays an icon of stacked servers with a magnifying glass. The text repeats \"ARCHIVING & ACCESS\" and lists: \"MAINTAIN SHARED REPOSITORIES\", \"PROTECT SENSITIVE RECORDS\", \"SUPPORT RAPID RETRIEVAL\". The fourth panel, returning to a cyan theme, is titled \"EXCHANGE & ENGAGEMENT\" and features an icon of two human figures exchanging data, which corresponds to the first sub-image showing a cropped view of this specific icon. The text repeats \"EXCHANGE & ENGAGEMENT\" and lists: \"ENABLE REAL-TIME COLLABORATION\", \"CIRCULATE KEY INSIGHTS\", \"PROMOTE CONTINUOUS LEARNING\". Connecting the bottom of these four panels is a horizontal line with circular nodes, labeled \"INFORMATION LIFECYCLE\" in a small box above the right section of the line, indicating a directional progression from left to right. Above the third panel, a small tag reads \"LIVE DATA FEED\". At the very bottom of the slide, a horizontal banner with a glowing cyan border contains the word \"OUTCOMES\" in large white text on the left. To the right of this, a list of bullet points reads: \"FASTER, EVIDENCE-BASED CHOICES\", \"STRONGER INNOVATION PIPELINES\", \"LOWER PROCESS UNCERTAINTY\", \"HIGHER CROSS-TEAM EFFICIENCY\", \"LASTING ORGANIZATIONAL MEMORY\". The overall color palette is dominated by neon cyan, electric blue, and accents of orange, creating a sleek, modern aesthetic.", "input_images": ["images/multi_image/afbbaa54-c08b-4ad0-a32e-254b0ca6e80a_input_0.jpg"], "target_image": "images/multi_image/afbbaa54-c08b-4ad0-a32e-254b0ca6e80a_target.png", "width": 1376, "height": 768}
+{"index": 280, "id": "db1fca74-3fda-4223-9c2d-94fc2c6472bb", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Emphasizing the concept of data‑driven urban innovation, this presentation slide uses a minimalist white background containing a central technology diagram, four rectangular illustration panels, and explanatory text sections. The visual language relies on black typography for clarity, paired with a bright cyan blue used consistently for headings, icons, connectors, and accent graphics.\n\nIn the upper left area, a prominent black sans‑serif title reads “URBAN INTELLIGENCE NETWORKS: BUILDING THE DIGITAL CITY,” bordered on its left side by a vertical cyan bar. Directly below, a smaller subtitle in regular black text states, “An integrated framework linking devices, platforms, and public services.”\n\nThe composition centers on a schematic illustrating information flow. On the left of this diagram, a globe‑and‑nodes symbol is captioned “INTERNET OF THINGS (IoT)” in bold black letters. In the middle, a cloud icon carries the label “CLOUD COMPUTING.” To the right, a bar‑chart symbol is marked “BIG DATA ANALYTICS.” Above the cloud graphic, the phrase “5G/6G NETWORKS” appears above a network‑node icon. These components are connected by thin black arrows: one arrow leads from “INTERNET OF THINGS (IoT)” to the “5G/6G NETWORKS” icon; another arrow continues from “5G/6G NETWORKS” to the “CLOUD COMPUTING” cloud. A two‑way arrow joins “CLOUD COMPUTING” and “BIG DATA ANALYTICS.” Additional arrows extend outward from the IoT and Cloud Computing elements toward the four surrounding illustration panels, indicating how the core platform supports multiple city services.\n\nAround this central graphic, four illustrated panels present different application domains, each with a descriptive heading and bullet list.\n\nIn the top‑left quadrant, the first image depicts a smiling woman inside a train looking at a holographic map. Beneath this panel, a bold header reads “CONTEXT‑AWARE TRAVEL SERVICES.” The bullet list underneath states: “– Live Journey Optimization,” “– Cross‑network Trip Coordination,” and “– Unified Ticketing & Billing.”\n\nIn the top‑right quadrant, the third image shows a couple walking on a path lined with streetlights. To the right of this illustration, the bold header reads “ADAPTIVE PUBLIC SPACES.” The associated bullet points read: “– Responsive Lighting Systems,” “– Urban Climate Monitoring,” and “– Automated Facility Control.”\n\nIn the bottom‑left quadrant, the second image shows a man riding a bicycle down a city street while looking at a phone. Below this panel, the bold header reads “ACTIVE & SHARED MOBILITY.” Its bullet list specifies: “– Connected Micromobility Services,” “– Low‑carbon Transport Options,” and “– Networked Charging & Parking.”\n\nIn the bottom‑right quadrant, the fourth image (not separately provided) shows the interior of a bus filled with diverse, smiling passengers. Beneath this illustration, the bold header states “CITIZEN‑FOCUSED DIGITAL TOOLS.” The supporting bullet points read: “– Tailored Travel Recommendations,” “– Inclusive Design & Assistance Features,” and “– Instant Service Alerts & Updates.”\n\nCyan square outlines appear decoratively in the upper‑right and lower‑left corners of the slide, while subtle cyan and pink glowing dots cluster near the bottom edges, adding a sense of motion and technology. The overall structure is balanced and orderly, with the central network diagram visually linking the four thematic pillars of the digital city ecosystem.", "input_images": ["images/multi_image/db1fca74-3fda-4223-9c2d-94fc2c6472bb_input_0.jpg", "images/multi_image/db1fca74-3fda-4223-9c2d-94fc2c6472bb_input_1.jpg", "images/multi_image/db1fca74-3fda-4223-9c2d-94fc2c6472bb_input_2.jpg", "images/multi_image/db1fca74-3fda-4223-9c2d-94fc2c6472bb_input_3.jpg"], "target_image": "images/multi_image/db1fca74-3fda-4223-9c2d-94fc2c6472bb_target.png", "width": 1376, "height": 768}
+{"index": 281, "id": "68b08fe0-e631-4af6-be08-3552a404cfd9", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are an AI slide-generation engine. Convert the supplied slide JSON into a high-quality presentation image.\n\nFollow these constraints exactly:\n- The JSON is authoritative. Render only the contents of `blocks`; never add new words.\n- Embed all source images into the slide.\n- `global_style` defines the shared visual identity. Preserve its fonts, colors, background, and style language.\n- `visual_hierarchy` describes what should feel dominant, secondary, or subtle.\n- `special_instructions` may refine or override the default look; respect it when present.\n\nDesign with variety. Do not default to a title plus bullet list. Use asymmetry, columns, modular cards, large focal visuals, restrained annotations, or generous negative space when appropriate. The final slide must feel clean, readable, and presentation-ready.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Multimodal Reasoning Dashboard\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#0C1220\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #22314F 0%, #0C1220 55%, #050710 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Montserrat\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#4DD0E1\\\",\\n      \\\"secondary\\\": \\\"#FFB74D\\\",\\n      \\\"text\\\": \\\"#F5F7FA\\\",\\n      \\\"accent\\\": \\\"#9FA8DA\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall feel is sleek and technical, echoing a modern research slide. Dark, slightly bluish background makes the colorful diagram pop. Thin neon-cyan lines and soft glows emphasize a high-tech, AI-themed atmosphere. Corners are subtly rounded; shadows are soft and diffused. Minimal clutter so the central diagram remains dominant.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Futuristic\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 1,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Visual-Text Reasoning Pipeline\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#F5F7FA\\\",\\n            \\\"font_family\\\": \\\"Montserrat\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, slightly above the image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Uppercase tracking slightly expanded; centered with generous letter spacing. Subtle cyan glow behind text to tie into the primary accent color.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/68b08fe0-e631-4af6-be08-3552a404cfd9_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/68b08fe0-e631-4af6-be08-3552a404cfd9_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Diagram showing an image-question reasoning workflow with diversification, rationalization, and ensemble stages.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning about 75% of slide width and 60% of height\\\",\\n            \\\"image_style\\\": \\\"clean with subtle drop shadow\\\",\\n            \\\"container\\\": \\\"rounded-rectangle panel with faint border\\\",\\n            \\\"special_instructions\\\": \\\"Panel has a very dark blue (#11182A) background with 12px rounded corners and a thin #4DD0E1 border at 40% opacity. The image itself fills most of this panel with 8px inner padding and slightly rounded corners. Apply soft ambient shadow to lift it from the background and keep it clearly dominant.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-1-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"From image and question to diversified answers, rationalized evidence, and a final ensemble choice.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#CFD8DC\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image panel\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Small, single-line text centered under the main diagram; keep line length short for quick readability.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-1-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Three semi-transparent gradient rings arranged behind the image panel, subtly hinting at the three stages of the pipeline.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette with cyan, amber, and soft violet\\\",\\n            \\\"position\\\": \\\"layered behind the central image, slightly offset left, right, and bottom\\\",\\n            \\\"special_instructions\\\": \\\"Rings have low opacity (~0.25), large radius, and soft blur, creating a halo effect without overpowering the main diagram.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nNow produce the final 16:9 slide image. Do not add extra text; prioritize clarity, hierarchy, and polished visual design.", "input_images": ["images/multi_image/68b08fe0-e631-4af6-be08-3552a404cfd9_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 282, "id": "42a2b73c-214a-43eb-a980-ac1246dc21af", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Design a slide titled \"CHOOSING THE RIGHT STRATEGIC SUPPLIER\" displayed in large, dark blue capital letters on a light blue background. On the left side, place two beige text boxes communicating the refreshed message: \"Operational reliability, technical depth, and partner experience shape the strongest suppliers.\" and \"Enduring collaboration and network reach drive continuous improvement and new opportunities.\" To the right, use a radar chart to present five updated attributes: \"Reliability,\" \"Product Excellence,\" \"Partner Experience,\" \"Strategic Impact,\" and \"Network Reach.\" Preserve the existing blue-to-green gradient shape on the radar chart, indicating consistently strong performance, with the points corresponding to \"Partner Experience\" and \"Reliability\" extending to the outer rings. Keep the surrounding illustrative icons aligned with each metric, and position a clasped-hands icon above the reliability-related label to emphasize collaboration. The first sub-image showing the close-up of the chart remains centered toward the right side of the overall slide layout.", "input_images": ["images/multi_image/42a2b73c-214a-43eb-a980-ac1246dc21af_input_0.jpg"], "target_image": "images/multi_image/42a2b73c-214a-43eb-a980-ac1246dc21af_target.png", "width": 1376, "height": 768}
+{"index": 283, "id": "f4390488-1997-4fa0-9b03-11d1aa27711e", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The presentation slide is set against a rich, deep brown background that features a soft, abstract gradient of glowing orange and pale blue light sweeping diagonally from the bottom left to the top right, creating a sense of warmth and clarity. Dominating the top of the slide in large, bold, white sans-serif capital letters is the title, \"ETHICAL INNOVATION: Curiosity, Responsibility, and Impact\". Directly beneath this, in a slightly smaller white serif font that is italicized, is the subtitle: \"Turning Bold Ideas into Meaningful, Human-Centered Change.\"\n\nThe body of the slide is organized into four distinct sections, arranged in a two-by-two grid, each introduced by a line-art icon in a muted gold color. In the top-left quadrant, next to an icon of a magnifying glass examining gears, the bold white header reads, \"FOUNDATION:\". Below this, a white bullet point defines the concept: \"A mindset that combines exploration, critical thinking, and empathy to design solutions that benefit people and the planet.\"\n\nMoving to the top-right quadrant, an icon of a hand holding a checkmark introduces the bold white header, \"REAL‑WORLD APPLICATIONS:\". Three white bullet points follow, listing practical applications: \"Developing technologies that prioritize user safety and dignity.\", \"Designing cities and spaces that are inclusive, accessible, and sustainable.\", and \"Creating business models that balance profit with social and environmental value.\"\n\nIn the bottom-left quadrant, a shield icon containing a key sits to the left of the bold white header, \"GUIDING VALUES:\". This section details three guiding concepts with white bullet points: \"Curiosity: Asking better questions and challenging assumptions.\", \"Integrity: Aligning actions, data, and decisions with clear ethical standards.\", and \"Stewardship: Considering long-term consequences for communities and ecosystems.\"\n\nFinally, the bottom-right quadrant features an icon of a gavel resting on a book next to the bold white header, \"PRACTICAL CONSTRAINTS:\". Two white bullet points clarify the boundaries of this approach: \"Innovations must comply with legal, cultural, and safety requirements.\", and \"Projects should avoid harm, protect vulnerable groups, and respect privacy and consent.\"\n\nIn the bottom-right corner of the slide, the white \"TED\" logo is displayed, with the tagline \"A GLOBAL COMMUNITY OF CURIOUS MINDS.\" appearing in smaller white text beneath it.", "input_images": ["images/multi_image/f4390488-1997-4fa0-9b03-11d1aa27711e_input_0.jpg"], "target_image": "images/multi_image/f4390488-1997-4fa0-9b03-11d1aa27711e_target.png", "width": 1376, "height": 768}
+{"index": 284, "id": "bcfeda51-1897-49cd-91ad-4468baaed310", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### MISSION\nTransform the input Structured Content JSON into one elegant, high-fidelity presentation slide.\n\n### STRICT CONTENT CONTROL\nOnly the JSON may supply visible content. Do not add explanatory copy, generated bullets, fake labels, placeholder text, or decorative words. If text is sparse, preserve the sparse feeling.\n\n### STYLE CONTROL\nHonor `global_style` as the governing design system. Use the specified background, color palette, typography, and overall visual mood. Apply `special_instructions` when provided, especially for atmosphere, lighting, texture, or layout emphasis.\n\n### BLOCK HANDLING\nRender visible blocks according to their role and hierarchy. Use `script` only as hidden context for the slide's theme and emotional tone. Keep text blocks readable and avoid cramped paragraphs.\n\n### IMAGE INTEGRITY & PRESERVATION\nWhen you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n\n### LAYOUT EXPECTATION\nDesign a slide with deliberate composition. Vary structure across slides: some may be visual-led, some text-led, some diagrammatic, some minimal. The output should feel premium, consistent, and ready for a professional deck.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Model Benchmark Highlight\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#0B1020\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #1B2640 0%, #050814 70%, #020309 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"IBM Plex Sans\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#FF6B35\\\",\\n      \\\"secondary\\\": \\\"#4ECDC4\\\",\\n      \\\"text\\\": \\\"#F7FAFC\\\",\\n      \\\"accent\\\": \\\"#F9C74F\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall feel is professional and data-driven, with a dark analytic dashboard vibe that makes the pale table background pop. Clean sans-serif typography, subtle glows, and thin neon accent lines reinforce a technical research atmosphere. Avoid clutter: let the table be the visual focus, with minimal text and a couple of soft gradient shapes in the corners to frame the content.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Analytical, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 20,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-20\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Gemma2-9B Retrieval Strategies: EM & ACC Overview\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#F7FAFC\\\",\\n            \\\"font_family\\\": \\\"IBM Plex Sans\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, above the main image with generous margin\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\\"shadow\\\"],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Large uppercase and title case mix, letter-spacing slightly increased. Subtle soft glow behind text using the primary orange accent to make it stand out against the dark gradient.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/bcfeda51-1897-49cd-91ad-4468baaed310_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/bcfeda51-1897-49cd-91ad-4468baaed310_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Performance table comparing retrieval-based methods on in-domain and out-of-domain QA benchmarks.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center of slide, spanning about 80% of width and 55% of height\\\",\\n            \\\"image_style\\\": \\\"subtle drop shadow with thin accent border\\\",\\n            \\\"container\\\": \\\"rounded rectangle frame with faint teal outline\\\",\\n            \\\"special_instructions\\\": \\\"Use a 12px border-radius on the image corners. Add a 2px border in #4ECDC4 and a soft outer glow in #0F172A to lift it from the dark background. Ensure it remains the dominant element with clear legible text.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-20-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Skill-RAG delivers the best average EM and ACC, especially on out-of-domain datasets.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#E2E8F0\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"just below the image, centered horizontally\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Small, concise line in regular weight, aligned center. Key phrase 'Skill-RAG' highlighted in the primary orange color within the text.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"caption-20-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"caption\\\",\\n          \\\"content\\\": \\\"Table: Exact Match (EM) and Accuracy (ACC) scores across HotpotQA, NQ, TriviaQA, MuSiQue, and 2Wiki benchmarks.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#A0AEC0\\\",\\n            \\\"position\\\": \\\"bottom-center, near slide footer\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"light\\\",\\n              \\\"text_effect\\\": [],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"none\\\"\\n            },\\n            \\\"special_instructions\\\": \\\"Small caption text, left-aligned within an invisible center-aligned text box, providing context without drawing attention away from the table.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-20-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A large, soft teal gradient ring partially visible behind the right side of the table, suggesting a data halo.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"right-center, behind the image layer\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (~18%), blurred edges, radius roughly 60% of slide height. Positioned so it peeks out from behind the top-right and bottom-right corners of the table.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-20-2\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A thin diagonal orange line with small data-point dots, running subtly from the bottom-left toward the center under the image.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"decorative background shape\\\",\\n            \\\"color_scheme\\\": \\\"primary and accent colors\\\",\\n            \\\"position\\\": \\\"bottom-left quadrant, beneath all text and image layers\\\",\\n            \\\"special_instructions\\\": \\\"Line should be subtle and semi-transparent (#FF6B35 at 35% opacity) with 4–5 tiny circular nodes in #F9C74F, evoking an upward trend without distracting from the table.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nProduce a professional 16:9 slide image that faithfully reflects the structured JSON and avoids any invented content.", "input_images": ["images/multi_image/bcfeda51-1897-49cd-91ad-4468baaed310_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 285, "id": "d7c264f5-502a-472f-880f-a14572d01d1d", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Designed in a calm, contemporary style dominated by soft teal and warm beige tones, this presentation slide features a wide white central canvas framed by a slim teal border. The slide’s title, centered at the top in a clean, dark sans‑serif font, reads: “Cozy Laundry Corner,” with a compact subtitle just below it stating: “Simple, Fresh, Organized.”\n\nIn the upper left corner, a small circular accent icon—a minimalist outline of a washing machine in teal—adds a subtle visual cue to the theme without drawing attention away from the main imagery.\n\nThe focal point of the slide is the sub image, placed prominently in the middle of the layout and occupying most of the horizontal space. The sub image shows a bright, stylish laundry area: a white front‑loading washing machine stands slightly left of center, its glass door tinted blue, with neatly folded knit blankets stacked on top. Textured surfaces and natural elements, such as a woven basket with a green plant, a teal laundry basket on the floor holding patterned textiles, and a white chair with a woven basket full of colorful folded clothes, create a cozy and organized atmosphere. A striped rug in warm browns stretches along the bottom, visually grounding the scene.\n\nBeneath the sub image, a soft beige band runs across the slide, containing two short content zones separated by a thin vertical teal line:\n\nOn the left side of this band is a solid teal label bar with small white text reading “Laundry Mood.” Directly under it, three brief bullet points in dark gray state:\n\n• Light, airy workspace  \n• Natural textures & soft colors  \n• Everything within easy reach  \n\nOn the right side, a matching teal label bar carries the heading “Daily Routine.” Below it, three concise bullets mirror the minimal style:\n\n• Sort, wash, fold  \n• Keep supplies visible  \n• Enjoy a tidy corner  \n\nAll supporting text is kept intentionally minimal, allowing the sub image to remain the dominant element. Generous white space around the photo and text blocks keeps the layout uncluttered, guiding the viewer’s eye naturally from the title down to the sub image and then to the short notes that gently reinforce the serene, well‑ordered laundry environment depicted in the scene.", "input_images": ["images/multi_image/d7c264f5-502a-472f-880f-a14572d01d1d_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 286, "id": "1a336909-e5b7-419e-8497-1321fcc49730", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "On the left, a large blue title reads **\"Efficient Gradient Sharing: E2E vs GRAM.\"** Beneath it, two rounded rectangular boxes contain brief text: the top white box states, *\"User interactions are encoded into content gradients for recommendation models,\"* while the lower blue box in light green text notes, *\"GRAM reduces redundant gradients by grouping shared content across users.\"*  \n\nTo the right, the main sub image is prominently displayed, showing a mini-batch of users and items at the top, followed by two parallel diagrams labeled **E2E** and **GRAM**. The E2E section depicts individual user–item pairs generating multiple gradients for the content encoder, while the GRAM section visually clusters shared items, resulting in fewer gradients flowing into the encoder.", "input_images": ["images/multi_image/1a336909-e5b7-419e-8497-1321fcc49730_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 287, "id": "037967b2-faff-4163-be98-dc054df2a8a5", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The presentation slide, titled \"NAVIGATING INDIA'S GOVERNANCE FRAMEWORK\" in large, bold, teal sans-serif lettering with a white outline set against a light teal banner at the top, features a deep blue background with a subtle, darker blue circuit-board-like line pattern at the bottom. On the left side, two white rounded rectangular text boxes with teal drop shadows contain key messages: the top box reads \"Policy initiatives seek inclusive growth, yet outcomes depend on coordination across multiple tiers of government.\" and the bottom box states \"Shifts in voter behavior and party alliances highlight the complexity of India's representative system.\" Both texts are in a clean, black sans-serif font.\n\nTo the right of the text boxes, the main visual element is a complex flowchart diagram enclosed in a light beige rounded rectangle. The diagram illustrates the relationship between different levels of government. At the top center of the diagram is a handshake icon labeled \"Tension\" below it. To the left is a box containing an icon of a government building with a flag, labeled \"CENTRAL GOVERNMENT (Union List)\" in bold black text. To the right is a larger box labeled \"STATE GOVERNMENTS (State List)\" in bold black text, containing various political party symbols (lotus, hand, elephant, etc.).\n\nBlue arrows connect these elements in a circular flow. From the \"CENTRAL GOVERNMENT\" box, an arrow points to a central box labeled \"Major Initiatives\" in bold black text. This box contains icons representing \"GST\", \"Digital India\", and construction equipment. Below these icons is the text \"Nationwide polls\". Another arrow points from \"Major Initiatives\" to the \"STATE GOVERNMENTS (State List)\" box. A third arrow points downwards from \"Major Initiatives\" to a box at the bottom labeled \"STATE GOVERNMENTS (State List)\", which also contains a government building icon. Finally, blue arrows loop back from the bottom \"STATE GOVERNMENTS\" box and the top-right \"STATE GOVERNMENTS (State List)\" box towards the top \"Tension\" handshake icon, completing the cycle. Additional small icons, including a handshake and scales of justice, are placed near the bottom-left corner of the diagram. The overall color scheme of the diagram uses blues, oranges, and greens for the icons and arrows.", "input_images": ["images/multi_image/037967b2-faff-4163-be98-dc054df2a8a5_input_0.jpg"], "target_image": "images/multi_image/037967b2-faff-4163-be98-dc054df2a8a5_target.png", "width": 1376, "height": 768}
+{"index": 288, "id": "b261307f-cf21-4d99-a47d-0293a144c0a6", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Dedicated to capturing a warm household moment, this presentation slide features a title banner reading **“FAMILY TIME: WORKING TOGETHER AT HOME”** at the top. The layout is divided into five vertical columns, each with a colored header, brief text, and a simple icon. The main sub image of a man and child cleaning a bathroom together is placed prominently across the lower center, spanning all columns as the main visual focus.\n\n**Column 1 – SHARED CHORES**  \n- Quick bathroom tidy-up  \n- Learning by doing together  \n\n**Column 2 – TEAMWORK**  \n- Divide simple tasks  \n- Support and encourage  \n\n**Column 3 – LIFE SKILLS**  \n- Build responsibility  \n- Practice everyday routines  \n\n**Column 4 – CONNECTION**  \n- Talk while you work  \n- Turn chores into moments  \n\n**Column 5 – CLEAN SPACE**  \n- Fresh, organized room  \n- Everyone helps maintain it", "input_images": ["images/multi_image/b261307f-cf21-4d99-a47d-0293a144c0a6_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 289, "id": "20f0941b-859d-4881-8e55-405b1aaacfac", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, titled **\"MEET THE MAJESTIC BEAR\"** in warm, rustic font across the top, uses a four‑column layout to highlight simple facts about this wild animal, with the **sub image** as the central visual focus.\n\n---\n\n**Column 1 – \"WHO IS THIS?\"**  \n- Large, centered **sub image** of the bear’s profile.  \n- Minimal caption:  \n  - \"A powerful wild bear\"  \n  - \"Thick fur, keen senses\"\n\n---\n\n**Column 2 – \"KEY FEATURES\"**  \n- Short bullet phrases:  \n  - \"Strong jaw and snout\"  \n  - \"Rounded ears\"  \n  - \"Dense brown coat\"\n\n---\n\n**Column 3 – \"HABITAT & LIFE\"**  \n- Minimal text beneath a small repeat of the **sub image** (cropped if needed):  \n  - \"Forests, mountains, riversides\"  \n  - \"Omnivore: plants, fish, small animals\"  \n  - \"Often solitary\"\n\n---\n\n**Column 4 – \"RESPECT & PROTECT\"**  \n- Short reminders:  \n  - \"Observe from a safe distance\"  \n  - \"Do not feed wild bears\"  \n  - \"Protect natural habitats\"", "input_images": ["images/multi_image/20f0941b-859d-4881-8e55-405b1aaacfac_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 290, "id": "d9847e7e-1f09-4a25-a046-c70a1aa40cb6", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are an elite AI Poster Composition Engine, fusing the imagination of a children’s book illustrator with the layout discipline of an information designer.\n\n**Your Goal:** Transform the incoming **Structured Content JSON** into a single, polished educational poster image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object describing ONE poster. Interpret it as follows:\n1. **`global_style`**: Establishes the fixed visual universe (lighting, surface, base palette).\n2. **`visual_hierarchy`**: Specifies the compositional strategy and emotional tone.\n3. **`blocks`**: The concrete elements (text, images, decorative illustrations) that must appear.\n\n### CORE RENDERING DIRECTIVES (NON-NEGOTIABLE)\n\n#### 1. IMAGE FIDELITY & PRESERVATION (TOP PRIORITY)\n* **Directive:** For every block with `type: \"image\"`, embed the referenced source image exactly as provided.\n* **Source Assets:** Images are given as attachments with this prompt. These exact files must be used.\n* **Hard Constraints:**\n  * Do **NOT** crop, stretch, repaint, or regenerate the content inside these images.\n  * Do **NOT** substitute charts, illustrations, or icons with new artwork.\n  * Do **NOT** omit any referenced image—each one must be present in the final poster.\n* **Permitted Enhancements:** You may stylize the *frame* or container (e.g., soft glow, stroke, subtle tilt, inset shadow) as long as the original pixels remain untouched and fully legible.\n* **Multiple Assets:** When more than one image is specified, you are responsible for clear, organized placement (rows, columns, or as governed by `special_instructions`).\n\n#### 2. TEXT SOURCING & LIMITS\n* **Directive:** The JSON `content` fields are the ONLY permitted source of textual material.\n* **Hard Constraints:**\n  * **No Invention:** Do not fabricate new phrases, labels, or taglines to fill gaps.\n  * **No Padding:** If there is very little text, keep the layout airy and premium; rely on whitespace instead of filler.\n  * **Formatting Control:** You may tune typographic details—line breaks, weight, letter spacing, and case—to optimize readability and rhythm, while avoiding dense text blocks.\n* **`role: \"script\"` Handling:** Any block labeled `script` is for your understanding only. Use it to inform ambience and lighting, but do **NOT** place that text on the poster unless explicitly instructed by `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" PRIORITY CHANNEL\n* **Directive:** The `special_instructions` key (in `global_style` or within a block) overrides standard behaviors whenever present.\n* **Behavior:**\n  * If a conflict arises (e.g., neutral background vs. “turn this into a starry night sky”), follow `special_instructions`.\n  * Leverage this field to introduce atmosphere—textures, glows, particle effects, stylized shadows—while respecting all other non-negotiable rules.\n\n#### 4. GLOBAL_STYLE CONSISTENCY RULESET\n* **Directive:** Treat `global_style` as the canonical visual system that must remain stable across independently rendered posters.\n* **Background:** Use the precise `background.color` value. The base background appearance must be identical for all posters sharing this style.\n* **Typography:** Apply `font_palette.primary` to main headings and `font_palette.secondary` to body and supporting text.\n* **Color Usage:** When a block references `color_ref`, map it directly to the matching hex value in `color_palette`.\n* **Role-Based Sizing:**\n  * TITLE elements (`importance: \"dominant\"`) must share identical scale and anchor position across all posters.\n  * BODY/LIST elements (`importance: \"supporting\"`) must retain consistent font size, line height, and styling.\n  * CAPTION elements must also be consistent from one poster to the next.\n* **Purpose:** This ensures multiple independently rendered posters feel like a cohesive series.\n\n#### 5. LAYOUT VARIETY & COMPOSITIONAL INTEREST\n* **Directive:** Avoid repetitive “headline plus bullet list” structures. Each poster should feel like a unique, thoughtfully designed layout.\n* **Implementation:**\n  * **Text-forward layouts:** Use columns, offset blocks, or staggered groupings rather than stacking text in a single vertical column.\n  * **Image-centric layouts:** Let illustrations or photos dominate when present; position any text as light, supportive labeling.\n  * **Hybrid layouts:** Seek a balanced relationship between imagery and text via side-by-side modules, overlapping panels, or layered cards.\n  * **Whitespace Discipline:** Treat empty space as a design asset. Do not artificially enlarge or duplicate content to fill the canvas.\n* **Variation Techniques:**\n  * Shift text alignment and anchoring (left, right, center, or asymmetric).\n  * Employ diverse container shapes (tags, bubbles, ribbons) where stylistically appropriate.\n  * Vary the perceived weight of regions—sometimes text-led, sometimes visual-led.\n* **Goal:** Produce an engaging set of posters where each composition feels fresh while still honoring the underlying system.\n\n### OUTPUT GENERATION\nRender a single, high-resolution poster image according to these specifications.\n\n### POSTER JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Animal Crossing Whimsical Astronomy\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Rounded MT Bold\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#E63972\",\n      \"secondary\": \"#8B4513\",\n      \"text\": \"#1A3C28\",\n      \"accent\": \"#FFD700\"\n    },\n    \"special_instructions\": \"The poster embraces a cheerful, game-inspired aesthetic reminiscent of Animal Crossing. Use a bright white canvas framed by a soft rainbow border. Typography is rounded, approachable, and optimized for kids. Character art should appear crisp with gentle shading and clean outlines.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Playful, Whimsical, Educational, Friendly\",\n    \"composition_style\": \"Grid-based, Three-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 771,\n      \"blocks\": [\n        {\n          \"id\": \"title-771\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"STAR-GAZING WITH FRIENDS\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Rounded MT Bold\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"outline\",\n                \"rainbow_fill\"\n              ],\n              \"outline_color\": \"#1A3C28\",\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Title characters use a smooth rainbow gradient (red to violet) with a dark green outline. Letterforms are chunky, rounded, and slightly spaced for a playful banner feel.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_78_139_408_436.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_78_139_408_436.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of owl character with telescope\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-left column\",\n            \"image_style\": \"cartoon illustration with white background\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Shows a brown owl character looking through a brass telescope on a tripod, surrounded by stars and a moon.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_531_144_846_437.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_531_144_846_437.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of pink owl character with telescope and star chart\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-center column\",\n            \"image_style\": \"cartoon illustration with white background\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Shows a pink owl character holding a star wand, standing next to a telescope and a star chart poster.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_979_134_1315_441.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_979_134_1315_441.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of human character pointing at shooting star\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-right column\",\n            \"image_style\": \"cartoon illustration with white background\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Shows a pink-haired human character pointing at a shooting star, with a telescope on a tripod nearby.\"\n          }\n        },\n        {\n          \"id\": \"body-771-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"GET SET TO OBSERVE\\n- Adjust the tripod and focus\\n- Keep both eyes relaxed\\n- Notice how bright each star is\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A3C28\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Section heading is bold and uppercase. Bullets use hyphens and are left-aligned for easy reading.\"\n          }\n        },\n        {\n          \"id\": \"body-771-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"MAP THE NIGHT SKY\\n- Match patterns to a star chart\\n- Write down what you discover\\n- Mark favorite constellations\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A3C28\",\n            \"importance\": \"supporting\",\n            \"position\": \"center column, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Section heading is bold and uppercase. Bullets use hyphens and are left-aligned for easy reading.\"\n          }\n        },\n        {\n          \"id\": \"body-771-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"CHASE NIGHTTIME WONDERS\\n- Watch for streaking meteors\\n- Spot colorful planet pairs\\n- Make a wish on bright trails\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A3C28\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Section heading is bold and uppercase. Bullets use hyphens and are left-aligned for easy reading.\"\n          }\n        },\n        {\n          \"id\": \"illustration-771-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A thin, rounded rectangular border surrounding the entire slide content with a rainbow gradient color scheme\",\n          \"style\": {\n            \"illustration_style\": \"decorative border\",\n            \"color_scheme\": \"rainbow gradient (red, orange, yellow, green, blue, purple)\",\n            \"position\": \"outer edge of slide\",\n            \"special_instructions\": \"The border has rounded corners and follows the rainbow color sequence smoothly.\"\n          }\n        },\n        {\n          \"id\": \"illustration-771-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Two thin vertical lines separating the three columns of content\",\n          \"style\": {\n            \"illustration_style\": \"divider lines\",\n            \"color_scheme\": \"light green\",\n            \"position\": \"between columns\",\n            \"special_instructions\": \"The lines are subtle and extend from below the images to the bottom of the text area.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nRender the final poster in a widescreen 16:9 aspect ratio with sharp, legible typography and a well-balanced, kid-friendly composition.\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the poster.**", "input_images": ["images/multi_image/d9847e7e-1f09-4a25-a046-c70a1aa40cb6_input_0.jpg", "images/multi_image/d9847e7e-1f09-4a25-a046-c70a1aa40cb6_input_1.jpg", "images/multi_image/d9847e7e-1f09-4a25-a046-c70a1aa40cb6_input_2.jpg"], "target_image": "images/multi_image/d9847e7e-1f09-4a25-a046-c70a1aa40cb6_target.png", "width": 1376, "height": 768}
+{"index": 291, "id": "29c4d63d-c1ee-4241-b0ed-05dcd6c9d37c", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This slide presents a focused **\"LAYOUT-GUIDED IMAGE GENERATION COMPARISON\"** with the subtitle **\"Visualizing Sub Image Variations Across Methods.\"**  \n\nThe layout is organized into two main rows. At the center of the slide, the **sub image** (the reference comparison grid featuring layouts, MIGC, 3DIS, and our method results) is prominently displayed, spanning most of the canvas width. The image illustrates object layouts in the leftmost column and generated results across different models in the remaining columns, for two scenarios: umbrellas with vases (top row) and colored sheep with a dog (bottom row).  \n\nAbove the sub image, a slim title bar runs horizontally, containing the main title on the left and the subtitle on the right, set against a subtle gradient background.  \n\nBelow the sub image, a simple three-column caption area summarizes the comparison with minimal text:\n- Left column titled **\"Layout Inputs\"** with a short note: *Object boxes & labels.*\n- Middle column titled **\"Baselines\"** with a note: *MIGC & 3DIS variations.*\n- Right column titled **\"Ours\"** with a note: *Depth & RGB fidelity.*  \n\nThe slide footer features a small line of text on the left reading **\"Layout-to-Image Generation Overview\"** and on the right **\"Method Comparison Snapshot\"**, both placed over a clean, light background to keep the attention on the sub image at the center.", "input_images": ["images/multi_image/29c4d63d-c1ee-4241-b0ed-05dcd6c9d37c_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 292, "id": "578115e0-8a6a-42e2-bef8-562377f0e8fc", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### MISSION\nTransform the input Structured Content JSON into one elegant, high-fidelity presentation slide.\n\n### STRICT CONTENT CONTROL\nOnly the JSON may supply visible content. Do not add explanatory copy, generated bullets, fake labels, placeholder text, or decorative words. If text is sparse, preserve the sparse feeling.\n\n### STYLE CONTROL\nHonor `global_style` as the governing design system. Use the specified background, color palette, typography, and overall visual mood. Apply `special_instructions` when provided, especially for atmosphere, lighting, texture, or layout emphasis.\n\n### BLOCK HANDLING\nRender visible blocks according to their role and hierarchy. Use `script` only as hidden context for the slide's theme and emotional tone. Keep text blocks readable and avoid cramped paragraphs.\n\n### IMAGE INTEGRITY & PRESERVATION\nWhen you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n\n### LAYOUT EXPECTATION\nDesign a slide with deliberate composition. Vary structure across slides: some may be visual-led, some text-led, some diagrammatic, some minimal. The output should feel premium, consistent, and ready for a professional deck.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Soft Couture Portrait\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#F9F4F5\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #FDF9FA 0%, #F0E6EF 55%, #E3D4E3 100%)\\\",\\n      \\\"texture\\\": \\\"paper\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Didot\\\",\\n      \\\"secondary\\\": \\\"Montserrat\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#D59BB5\\\",\\n      \\\"secondary\\\": \\\"#9F7FAF\\\",\\n      \\\"text\\\": \\\"#3C3340\\\",\\n      \\\"accent\\\": \\\"#F7C7D9\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall look is airy, romantic, and editorial, inspired by high-fashion magazines. Use lots of negative space and soft edges. Background has a subtle paper grain to keep the pastel palette from feeling flat. Corners are slightly rounded; shadows are soft and diffused as if lit by a studio softbox. No harsh lines—favor gentle gradients and delicate strokes that echo the puff sleeves and pastel pink tones of the dress.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Elegant, Dreamy, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 15,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-15\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Whispers of Blush\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#3C3340\\\",\\n            \\\"font_family\\\": \\\"Didot\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with generous margin above the image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Large, elegant serif set in all caps with slightly expanded letter-spacing; soft drop shadow in #E1D5E4 to lift it from the gradient background.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/578115e0-8a6a-42e2-bef8-562377f0e8fc_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/578115e0-8a6a-42e2-bef8-562377f0e8fc_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Soft-focus portrait of a person in a pale pink, open-back dress with voluminous sleeves against a muted blue background.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, occupying about 70% of slide height and 60% of width, directly beneath the title\\\",\\n            \\\"image_style\\\": \\\"no frame, subtle outer glow\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Apply a gentle rounded-corner mask (18px radius) to the image with a very soft, diffused shadow (#BDAAC9 at 25% opacity) to emphasize depth while keeping the photo as the dominant focal point.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-15-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"A study in weightless volume and pastel serenity.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Montserrat\\\",\\n            \\\"color\\\": \\\"#6A5A72\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, aligned just below the image\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Small, uppercase sans-serif with moderate letter-spacing, center-aligned to keep the layout balanced and understated.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-15-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A large, translucent circle overlapping the back of the image, echoing the pale blue background of the photo to create a halo effect.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette, blending #B5C6E0 and #F7C7D9 at low opacity\\\",\\n            \\\"position\\\": \\\"centered behind the image, slightly offset to the right\\\",\\n            \\\"special_instructions\\\": \\\"Circle should be about 120% of the image width, with 30% opacity, soft edges, and layered behind the photo to subtly frame the subject without competing for attention.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate a refined 16:9 slide image suitable for a professional deck, with clear hierarchy, accurate content, and no unnecessary clutter.", "input_images": ["images/multi_image/578115e0-8a6a-42e2-bef8-562377f0e8fc_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 293, "id": "0ef61e84-4ee4-4fb8-86cb-dfb559314f62", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "With a sleek, minimalist white background, this slide highlights energy-aware evaluation under the heading **\"NAS ENERGY LANDSCAPE\"** in bold, dark-red text at the top.  \nAt the center, the provided **sub image** is prominently displayed, showing a gradient progression from **high energy tasks** on the left to **low energy tasks** on the right, emphasizing different NAS evaluation regimes.  \n\nBeneath the visual, a concise caption reads:  \n**\"From vanilla NAS to one/few-shot strategies: reducing evaluation cost while preserving performance insight.\"**", "input_images": ["images/multi_image/0ef61e84-4ee4-4fb8-86cb-dfb559314f62_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 294, "id": "cfb2a9ef-fb05-4467-9bfb-5964c742fa75", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Against a softly textured, light beige background that resembles aged paper, the slide opens with a bold, dark green title in large capital letters at the upper left:  \n“URBAN COURTSCAPE—MOMENT IN MOTION.”  \nA thick dark green underline extends beneath the title, echoing the graphic style of a design blueprint.\n\nBeneath the title, the left half of the slide is reserved for minimal text in two white rectangular boxes with angled right corners, floating one above the other:\n\n- The upper box contains a single, short line of black sans‑serif text:  \n  “Capturing a quiet pause beneath the hoop in an open-air court.”\n\n- The lower box, aligned just below, reads:  \n  “Contrast of school‑inspired attire and worn architecture frames the scene.”\n\nBoth boxes are slightly offset to the right to create a sense of layered cards resting atop the beige background. Small, scattered green dots decorate the left margin as subtle accents.\n\nOn the right half, dominating the slide, is the sub image. It is placed inside a thin rounded rectangle border in muted dark red, visually linked to the basketball backboard in the photograph. The sub image shows a person standing under a red outdoor basketball hoop structure. The individual wears a white long‑sleeved shirt, a vivid red neckerchief, a plaid skirt, and a black belt. Their face is intentionally blurred, keeping the focus on posture and clothing. Behind them, an aged building with wooden textures and a muted roofline stretches horizontally, while the overcast sky fills the background. The hoop and backboard tower above, angled diagonally from the top edge of the frame, creating a strong geometric line.\n\nBehind the sub image, partially overlapping it, a large, faint circular wash of soft green is placed, echoing the color accent used in the title underline. This circle lightly frames the upper right of the sub image, making the red of the backboard and neckerchief stand out. To anchor the layout, a thin dark line extends from the lower right corner of the sub image to the bottom edge of the slide, suggesting the continuation of the court structure just outside the frame.\n\nThe overall composition emphasizes the photograph as the central element—text is sparse and unobtrusive, while the muted background, circular accent, and angled linework draw the viewer’s eye back to the sub image and the still, contemplative moment it depicts.", "input_images": ["images/multi_image/cfb2a9ef-fb05-4467-9bfb-5964c742fa75_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 295, "id": "0b87770c-549d-4dfc-bcb3-ebbe134b3457", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide features a serene, traditional East Asian aesthetic with a soft, textured cream-colored background that resembles rice paper. Dominating the top center is the main title in a bold, dark blue serif font: \"NURTURING HARMONY: A FRAMEWORK FOR SHARED GROWTH.\" The slide is organized into four distinct vertical columns, each dedicated to a specific theme, separated by thin, elegant vertical lines that are interrupted by small decorative diamond motifs in the middle.\n\nThe first column on the far left is titled \"BUILDING CONNECTIONS & TRUST\" in a capitalized, dark blue sans-serif font. Below this heading is an illustration—corresponding to the first input image—depicting two magpie-like birds with black and white plumage perched on a blossoming branch with pink flowers, while three smaller birds fly in the background. Beneath this artwork, three bullet points in black serif text read: \"Creating welcoming gathering spaces,\" \"Encouraging meaningful conversations,\" and \"Celebrating shared traditions and stories.\"\n\nMoving to the second column, the header reads \"CARE, SUPPORT & RESILIENCE\" in the same capitalized style. The accompanying illustration, derived from the second input image, shows two elegant white cranes standing in shallow water near a rock, with tall reeds to the left, a pink lotus flower in the foreground, and a traditional pagoda-style pavilion in the background. The text below this image lists three bullet points: \"Checking in on one another regularly,\" \"Coordinating assistance during crises,\" and \"Promoting emotional and physical well-being.\"\n\nThe third column is labeled \"GROWTH, LEARNING & INSPIRATION.\" Its visual element, matching the third input image, portrays a scene with three swallows in flight, a blooming peony bush, an unrolled scroll with text at the bottom left, and a lantern hanging on the right. The text content for this section appears both above and below the image. Above the image, the bullet points are: \"Community classes and cultural programs,\" \"Sharing books, tools, and creative resources,\" and \"Guidance and role models for emerging leaders.\" Identical text is repeated below the image: \"Community classes and cultural programs,\" \"Sharing books, tools, and creative resources,\" and \"Guidance and role models for emerging leaders.\"\n\nThe final column on the right is titled \"CARING FOR OUR SHARED ENVIRONMENT.\" The illustration here, corresponding to the fourth input image, features a pond scene with three ducks swimming, purple iris plants on the left, a cluster of yellow chrysanthemums and other flowers on the right, and a large, porous rock formation. The text below this image lists three points: \"Collaborative clean-up and restoration days,\" \"Eco-friendly habits in homes and public spaces,\" and \"Protecting local water, wildlife, and green areas.\"\n\nDecorative elements enhance the traditional feel; subtle cloud-like line drawings appear in the top-left and top-right corners. Additionally, red vertical rectangular stamps, reminiscent of traditional Chinese seals, are placed in the top-right and bottom-right corners of the slide.", "input_images": ["images/multi_image/0b87770c-549d-4dfc-bcb3-ebbe134b3457_input_0.jpg", "images/multi_image/0b87770c-549d-4dfc-bcb3-ebbe134b3457_input_1.jpg", "images/multi_image/0b87770c-549d-4dfc-bcb3-ebbe134b3457_input_2.jpg", "images/multi_image/0b87770c-549d-4dfc-bcb3-ebbe134b3457_input_3.jpg"], "target_image": "images/multi_image/0b87770c-549d-4dfc-bcb3-ebbe134b3457_target.png", "width": 1376, "height": 768}
+{"index": 296, "id": "ec43752b-b580-4973-b842-b89bdc0addfc", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Creating this slide integrates a central illustration from the sub image, which showcases a futuristic, purple‑lit architectural tower rising above a modern cityscape at dusk. The slide design positions this tower against a clean, gradient background that transitions from soft orange to deep blue, echoing the sky in the sub image. A sleek dark banner at the top displays the title in white text: “CITY OF LIGHT: FUTURE SKYLINE”. Beneath the banner, the sub image occupies the majority of the slide, framed by a thin white border. Two small, unobtrusive text captions sit at the bottom corners in white font; the left reads “Innovation reshapes the urban horizon.” and the right reads “Architecture becomes a beacon after sunset.”", "input_images": ["images/multi_image/ec43752b-b580-4973-b842-b89bdc0addfc_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 297, "id": "ddea751b-3ef0-4a4a-84e5-4eb59b86d740", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, styled with a retro, comic-book aesthetic and blue headers, uses a Gundam robot motif to explain technical concepts. The main title at the top reads \"INTERACTIVE MECHA DYNAMICS\" in bold black text, with the subtitle \"A Gundam-Themed Tour of Real‑Time Motion\" directly below. A small blue banner in the top right corner says \"SIMULATION LAB SERIES\". The layout is divided into four quadrants. The top-left quadrant is titled \"UNDERSTANDING A MECHA UNIT\" and features a small version of the first input image showing a Gundam labeled with \"Parts,\" \"Shield,\" and \"Gundam.\" Next to it, the second input image (a red shield with arrows) is placed under the label \"External Influence (F),\" accompanied by text defining \"Rigid Chassis: Shape Preserved,\" \"Physical Attributes (Mass, Inertia Matrix),\" and \"Dynamic State (Pose, Linear & Angular Speed).\" The top-right quadrant, titled \"UPDATE LOOP OVERVIEW,\" illustrates a three-step process labeled \"1. Scene Setup (Models, Materials, Inputs),\" \"2. Per-Frame Advancement (Integrate Motion),\" and \"3. World Effects (Gravity, Contact, Drag),\" which flows into a box labeled \"Detection, Response & Stabilization.\" The bottom half features two sections: \"CONTACT DISCOVERY\" on the left, which displays the third input image depicting a battle scene with labels for \"Coarse Bounds (AABB),\" \"Discrete vs. Swept Tests,\" and \"Building the Contact Set\"; and \"SOLVER CORE & JOINT CONTROL\" on the right. This final section includes the fourth input image showing a mechanical arm with labels \"Arms,\" \"Arm,\" \"Linkage,\" \"Linkages,\" \"Toms,\" and \"Hinges,\" alongside text reading \"Enforcing Limits and Joint Rules (No Overlap, Articulation)\" and \"Distributing Impulses for Stable, Believable Motion\" above a bar chart.", "input_images": ["images/multi_image/ddea751b-3ef0-4a4a-84e5-4eb59b86d740_input_0.jpg", "images/multi_image/ddea751b-3ef0-4a4a-84e5-4eb59b86d740_input_1.jpg", "images/multi_image/ddea751b-3ef0-4a4a-84e5-4eb59b86d740_input_2.jpg", "images/multi_image/ddea751b-3ef0-4a4a-84e5-4eb59b86d740_input_3.jpg"], "target_image": "images/multi_image/ddea751b-3ef0-4a4a-84e5-4eb59b86d740_target.png", "width": 1376, "height": 768}
+{"index": 298, "id": "bbdf0c45-8d9a-4f9a-a62d-d14b5905b0a3", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Design a slide set against a soft, light-blue background adorned with subtle decorative elements: a pale yellow sunburst in the upper left corner, faint swirling pastel lines in the upper right, and small white sparkles near the bottom left.  \n\nAt the top center, place a large, rounded rectangular banner with a pale yellow border and a light pink fill. Inside this banner, set the title in a bold, cream‑colored serif font with a soft drop shadow:  \n**“Finding the Trigger: Language Models at Work”**  \n\nBeneath this banner, position the input sub image as the main focal point, centered horizontally on the slide. Keep the sub image at a generous size so that its internal text and diagram elements remain clearly readable. Surround the sub image with a thin white border and a subtle drop shadow to make it feel like a mounted panel resting above the background.\n\nBehind the sub image, add two overlapping, softly rounded shapes in pastel tones — one pale yellow and one light blue — peeking out from behind the left and right edges of the image, respectively. These shapes echo the curved banner at the top and visually anchor the sub image in the middle of the slide.\n\nTo the left of the sub image, place a slim, vertical text box with rounded corners and a light‑blue fill. Give it a thin white border and a faint drop shadow. Inside, in a clean black sans‑serif font, write minimal guiding text:  \n- **Sentence, Event Type, Definition**  \n- **Model infers the trigger word**\n\nTo the right of the sub image, mirror this with another rounded, light‑blue text box of similar height but slightly wider. It also has a white border and soft shadow. Inside, in the same sans‑serif font, add:  \n- **Input → Generative Model**  \n- **Output: “The trigger is funneled.”**\n\nAt the bottom center of the slide, beneath the sub image and text boxes, include a small, rounded pill‑shaped label with a pale yellow fill and white border. In compact black sans‑serif text, write:  \n**Core Idea: Highlighting how models detect event triggers from context.**\n\nEnsure all elements are aligned so that the sub image dominates the composition, with the surrounding decorations and concise text serving only to frame and clarify its role as the central visual explanation.", "input_images": ["images/multi_image/bbdf0c45-8d9a-4f9a-a62d-d14b5905b0a3_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 299, "id": "543d35bf-1041-41f1-9614-a36ed604bb8f", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This slide presents an educational overview of **\"Cross‑Modal Alignment\"** using a clean white background and a layout divided into a central title and three distinct vertical columns below. At the top center, the words **\"Cross‑Modal Alignment\"** are displayed in a modern, bold font with a black outline; each word is filled with a gradient that shifts from teal to lime green, echoing the green overlays in the sub image and creating a precise, technical aesthetic.\n\nThe content is organized into three sections from left to right.  \n\n---\n\n### Left Section – **Imaging Domains**\n\n- Heading at the top: **\"Where It Appears\"** in bold, black, sans‑serif text.  \n- Directly beneath this heading, the sub image is inserted, scaled so that the entire collage of medical, satellite, aerial, and ground‑view examples is clearly visible.  \n- A thin, light‑gray frame surrounds the sub image to emphasize it as the focal visual.  \n- Below the sub image, a short, centered caption in small black text reads:  \n  **\"From medical scans to maps and street scenes, different sensors capture the same world.\"**\n\n---\n\n### Middle Section – **Shared Features**\n\n- Heading: **\"What Gets Aligned\"** in the same bold, black style.  \n- Under the heading, a slightly zoomed‑in duplicate of the sub image is placed, cropped to highlight the bright green correspondence lines linking paired images across the grid.  \n- Around this cropped view, subtle arrow icons point toward the green streaks, reinforcing the idea of correspondence.  \n- A brief line of text centered underneath states:  \n  **\"Common structures are matched across modalities using shared visual patterns.\"**\n\n---\n\n### Right Section – **Why It Matters**\n\n- Heading: **\"Why Alignment Helps\"** in bold, black type.  \n- Beneath, another focused crop of the sub image is used, this time emphasizing the bottom‑right examples (ground thermal–visible and driving scenes) to hint at real‑world decision making.  \n- To the right of this crop, three very short bullet points in a small, clean font read:\n  - **\"Fuse complementary data\"**  \n  - **\"Improve detection & mapping\"**  \n  - **\"Enable robust perception\"**  \n\nAt the bottom right corner of the entire slide, a small monochrome icon of a stylized sensor array (stacked rectangles with radiating lines) is placed, echoing the cross‑sensor theme in the sub image while keeping the slide visually minimal and letting the sub image remain the main focus.", "input_images": ["images/multi_image/543d35bf-1041-41f1-9614-a36ed604bb8f_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 300, "id": "84c9ae95-6373-4cb0-8492-1bc6da01c2f5", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide adopts a cool, dignified design inspired by the patina of aged copper, echoing the sub image. The background is a smooth vertical gradient that transitions from a muted twilight blue at the top to a deeper slate green at the bottom, giving the sense of an early evening sky. A faint texture resembling brushed metal adds depth while remaining subtle so the composition feels clean and modern.\n\nAt the very top center of the slide, a concise main title appears in a large, all-caps sans-serif typeface. The letters are rendered in a soft off‑white with a slight teal tint, creating a gentle contrast against the darker background. The title reads:  \n“STANDING TALL: SYMBOL OF HOPE”\n\nDirectly beneath the title, aligned to the center, is a short, single-line subtitle in a smaller, lighter-weight sans-serif font, also in off‑white. The subtitle text reads:  \n“An iconic figure illuminating the horizon”\n\nThe central focus of the slide is the sub image. It is placed prominently in the middle of the slide, occupying most of the vertical space so that it commands immediate attention. The sub image is framed within a thin, elegant border that echoes the oxidized copper color of the statue’s robes. The frame has gentle rounded corners and a subtle drop shadow, lifting the image slightly from the background and reinforcing it as the primary visual element.\n\nBehind the sub image, a wide, soft-edged halo of desaturated teal radiates outward, suggesting an atmospheric glow around the statue’s silhouette. This halo is very light and blends into the background gradient, creating a visual anchor without overwhelming the image.\n\nBelow the sub image, a narrow horizontal band stretches across the slide, aligned centrally and spanning roughly two-thirds of the width. This band is a semi-transparent slate rectangle, allowing the background gradient to show through. Within this band, three very short phrases are evenly spaced, each separated by a small vertical line in a muted bronze tone. The phrases are set in a minimalist sans-serif font and read, from left to right:\n\n- “Liberty”  \n- “Welcome”  \n- “Endurance”\n\nEach word is centered within its segment, with ample letter spacing, contributing to a clean and airy appearance. The bronze separators subtly echo the torch flame from the sub image, tying the palette together.\n\nIn the bottom-right corner of the slide, a small, simple icon of a torch is outlined in bronze. The flame is suggested by a single curved shape filled with a warm golden gradient. This icon is minimal and flat, staying in the background visually but reinforcing the theme represented in the sub image. Opposite this icon, in the bottom-left corner, a short, understated label appears in a small, all-caps font, in muted teal:  \n“GLOBAL LANDMARK PERSPECTIVE”\n\nThe slide maintains plenty of negative space around all textual elements, ensuring that the viewer’s eye is always drawn back to the sub image. No additional decorative patterns or complex graphics are used, allowing the statue in the sub image to remain the dominant feature and the central storytelling element of the design.", "input_images": ["images/multi_image/84c9ae95-6373-4cb0-8492-1bc6da01c2f5_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 301, "id": "0dfb6173-4473-4231-bd9b-d896e81e6b75", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Set against a background of soft, salmon-pink and white vertical stripes, the presentation slide features a structured infographic centered around the concept of \"PROFESSIONAL INTEGRITY.\" The title appears at the very top in a large, bold, black sans-serif font within a white rectangular header. Directly beneath this, a light orange banner spans the width of the content area, displaying the subheading \"GUIDING VALUES\" in black uppercase letters.\n\nThe main body of the slide is organized into three distinct vertical columns, each set against a teal background with black borders, outlining key dimensions of integrity.\n\nThe leftmost column is titled \"FAIR PRACTICES\" in bold black uppercase text. Below the title, small icons of a balance scale and a handshake—visual echoes of the first sub-image—sit above the text. The column’s content reads: \"Uphold fairness and respect in every interaction. Honor commitments and treat all stakeholders equitably. (e.g., Clear contracts, unbiased evaluations, respectful negotiation).\"\n\nThe center column focuses on \"SUSTAINABLE OPERATIONS,\" also written in bold black uppercase letters. This section features icons of a tree and a recycling symbol above the text. The description provided is: \"Design processes that preserve resources for the long term. Integrate sustainability into daily activities. (e.g., Efficient use of energy, circular use of materials, responsible sourcing).\"\n\nThe rightmost column, labeled \"STAKEHOLDER PARTNERSHIP\" in bold black uppercase, displays icons of a group of people and a heart. The text here states: \"Build meaningful collaboration with internal and external stakeholders. Encourage dialogue and shared value creation. (e.g., Feedback forums, joint initiatives, supporting partner development).\"\n\nOverlapping the bottom-right corner of the final column is a graphic of a magnifying glass with a brown handle, highlighting the word \"OUTCOMES\" in black text within the lens.", "input_images": ["images/multi_image/0dfb6173-4473-4231-bd9b-d896e81e6b75_input_0.jpg"], "target_image": "images/multi_image/0dfb6173-4473-4231-bd9b-d896e81e6b75_target.png", "width": 1376, "height": 768}
+{"index": 302, "id": "254d30db-2f7c-4dea-b706-c5cd9b04865e", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### ROLE\nYou are a senior presentation art director and precision layout renderer.\n\n### TASK\nRender the given Structured Content JSON as one polished 16:9 presentation slide. Treat the JSON as the complete design brief: `global_style` controls the visual system, `visual_hierarchy` guides composition, and `blocks` provide the exact slide elements.\n\n### HARD RULES\n1. Use only text found in block `content`. Do not invent titles, labels, captions, summaries, or decorative copy.\n2. Follow `global_style` strictly for background, palette, typography, lighting, spacing language, and overall material feel.\n3. If `special_instructions` appears, it is the creative override. Apply it faithfully while still preserving the provided textual content.\n4. Avoid generic slides. Choose a composition that matches the content: editorial grid, hero visual, split layout, card system, timeline, layered diagram, or spacious minimalist treatment.\n5. Embed all source images into the slide.\n\n### QUALITY BAR\nThe result should look like a professionally designed keynote slide: crisp text, intentional hierarchy, balanced whitespace, coherent color usage, and no clutter.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Tropical Travel & Adventure\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#8FD3E6\",\n      \"secondary\": \"#2C3E50\",\n      \"text\": \"#1A1A1A\",\n      \"accent\": \"#E8F4F8\"\n    },\n    \"special_instructions\": \"The slide carries a refined, contemporary tropical journey aesthetic. Maintain a split layout with a large, vibrant anime-style illustration on the left and organized informational cards on the right. Use tropical blues, teals, and hints of warm sunset tones to create an uplifting, aspirational mood. Typography remains bold, clear, and sans-serif for strong readability.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Playful, Inviting, Professional\",\n    \"composition_style\": \"Two-column (Image Left, Content Right)\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 571,\n      \"blocks\": [\n        {\n          \"id\": \"title-571\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"ISLAND HORIZONS GETAWAY\",\n          \"style\": {\n            \"color\": \"#FFFFFF\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center, above main heading\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"bubble\",\n            \"special_instructions\": \"Text is white, uppercase, centered inside a rounded rectangular pill shape with a light blue background (#8FD3E6).\"\n          }\n        },\n        {\n          \"id\": \"title-571-main\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"SUNLIT SHORES, ENDLESS JOURNEYS\",\n          \"style\": {\n            \"color\": \"#1A1A1A\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center, below subtitle\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, bold, uppercase black text centered on the white background.\"\n          }\n        },\n        {\n          \"id\": \"image-571\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_40_39_520_766.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_40_39_520_766.jpg\"\n              }\n            ],\n            \"source\": \"Main slide illustration\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left side, vertically spanning from top to bottom\",\n            \"image_style\": \"rounded corners\",\n            \"container\": \"none\",\n            \"special_instructions\": \"A vertical anime-style illustration showing a couple on a wooden deck overlooking a tropical ocean with overwater bungalows, coral reefs, and palm trees. The image has rounded corners.\"\n          }\n        },\n        {\n          \"id\": \"body-571-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"COASTAL HIDEAWAYS\\nSoft Sand Lagoon Beaches\\nPanoramic Ocean Villas\\nDrift-free Island Transfers\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, first card\",\n            \"container\": \"rounded panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a white rounded rectangle card with a light blue border. Title is bold uppercase. List items are normal weight. Card has a circular illustration of a bungalow on an island at the top.\"\n          }\n        },\n        {\n          \"id\": \"body-571-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SERENITY & RENEWAL\\nOcean Breeze Spa Rituals\\nGolden Hour Beachflow Yoga\\nCalm-Current Seaside Reflection\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, second card\",\n            \"container\": \"rounded panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a white rounded rectangle card with a light blue border. Title is bold uppercase. List items are normal weight. Card has a circular illustration of a person doing yoga under palm trees at the top.\"\n          }\n        },\n        {\n          \"id\": \"body-571-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"OCEAN DISCOVERY QUESTS\\nShimmering Reef Snorkel Trails\\nBlue Depth Dive Expeditions\\nClose-Up Encounters with Sea Life\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, third card\",\n            \"container\": \"rounded panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a white rounded rectangle card with a light blue border. Title is bold uppercase. List items are normal weight. Card has a circular illustration of underwater coral and fish at the top.\"\n          }\n        },\n        {\n          \"id\": \"body-571-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"LOCAL RHYTHMS & FLAVORS\\nHarbor Villages & Island Paths\\nSunset Dhoni Voyages & Cast-Line Fun\\nSea-to-Table Island Tastings\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, fourth card\",\n            \"container\": \"rounded panel\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a white rounded rectangle card with a light blue border. Title is bold uppercase. List items are normal weight. Card has a circular illustration of a boat at sunset at the top.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```", "input_images": ["images/multi_image/254d30db-2f7c-4dea-b706-c5cd9b04865e_input_0.jpg"], "target_image": "images/multi_image/254d30db-2f7c-4dea-b706-c5cd9b04865e_target.png", "width": 1376, "height": 768}
+{"index": 303, "id": "3499a688-57b7-4fda-b02f-886a0d72374e", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "The slide is titled **\"DATA-DRIVEN COLLABORATION\"** in large, dark grey capital letters at the top center, with the subtitle **\"INSIGHTS FOR A CONNECTED WORLD\"** below it, accompanied by a small tablet and globe icon. The background is white with a faint geometric network pattern in the corners. The content is divided into three columns.\n\nThe first column features the sub image (the two professionals reviewing charts on a tablet) above the heading **\"TEAM ANALYTICS\"** and the text:  \n\"Shared dashboards | Real-time updates | Clear visual reports.\"\n\nThe middle column displays a cropped view of the sub image focused on the tablet and chart icons above **\"DATA INSIGHTS\"** and the text:  \n\"Trends & forecasts | Performance metrics | Evidence-based decisions.\"\n\nThe third column shows a cropped view of the sub image focused on the globe graphic above **\"GLOBAL IMPACT\"** and the text:  \n\"Worldwide perspective | Market opportunities | Connected economies.\"\n\nA footer at the bottom right reads **\"VISUAL ANALYTICS SHOWCASE | PRESENTATION 2024\"**.", "input_images": ["images/multi_image/3499a688-57b7-4fda-b02f-886a0d72374e_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 304, "id": "f84baaae-907e-413c-a793-20f1cce54a30", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The presentation slide, titled \"CHANNEL SETTING IN JEWELRY\" in large, 3D-styled block letters resembling blue and yellow Lego bricks, introduces the concept as \"A Sleek River of Gemstones.\" The layout is divided into three main vertical sections against a white background, continuing the Lego-themed visuals to explain jewelry concepts.\n\nOn the left, the \"sub-image corresponding to the ring\" remains prominently displayed. This graphic shows a grey Lego ring structure mounted on a grey baseplate. The exterior of the ring is studded with translucent, round studs representing gemstones. Callout lines with black text identify specific components: \"CHANNEL-SET GEMSTONES (e.g., DIAMONDS)\" points to the studs; \"PARALLEL METAL WALLS\" indicates the structures that secure the stones in a groove; and \"METAL BAND (LEGO BRICKS)\" points to the grey band itself. Below this main illustration, a section titled \"WHAT IS CHANNEL SETTING?\" in bold black text provides an explanation using bullet points:\n* \"Stones are suspended between two parallel metal walls, forming a “channel.”\"\n* \"Gemstones sit edge-to-edge in a continuous line with minimal gaps.\"\n* \"Creates a smooth, modern look with an unbroken stream of sparkle.\"\n* \"Grooves in the metal walls hold the stones firmly without individual prongs.\"\nA circular magnified inset zooms in on the \"CHANNEL-SET GEMSTONES\" and \"PARALLEL METAL WALLS\" to show detail.\n\nIn the center column, the \"sub-image showing the Lego minifigure\" is placed. This illustration depicts a Lego construction worker minifigure with brown hair, a white shirt, and blue overalls, hammering a stud into a grey Lego brick. Below this figure is a section titled \"COMMON CHANNEL STYLES\" in bold, containing a list of three styles, each accompanied by a small Lego brick icon:\n* A blue brick icon next to \"STRAIGHT CHANNEL (Classic, linear rows of stones)\"\n* A white brick icon next to \"TAPERED CHANNEL (Stones graduate in size toward the center)\"\n* A grey brick icon next to \"DOUBLE CHANNEL (Two parallel rows for extra brilliance)\"\nA small, separate graphic of a minifigure pointing upwards is situated to the right of this text block.\n\nOn the right side, a light grey rectangular box is titled \"KEY FEATURES & BENEFITS\" in black text. This section lists five advantages, each marked by a distinct colored icon on the left:\n* A yellow star icon next to \"* CONTINUOUS SPARKLE: Stones line up to form a bright, uninterrupted row.\"\n* A blue brick icon next to \"* CLEAN PROFILE: No tall prongs, giving a low, snag-resistant silhouette.\"\n* A green brick icon next to \"* MODERN AESTHETIC: Ideal for sleek, streamlined ring designs.\"\n* A yellow brick icon next to \"* DESIGN FLEXIBILITY: Works well for bands, anniversary rings, and accents.\"\n* A blue and yellow shield icon next to \"* PRACTICAL DURABILITY: Protected stone edges reduce chipping risk.\"", "input_images": ["images/multi_image/f84baaae-907e-413c-a793-20f1cce54a30_input_0.jpg", "images/multi_image/f84baaae-907e-413c-a793-20f1cce54a30_input_1.jpg"], "target_image": "images/multi_image/f84baaae-907e-413c-a793-20f1cce54a30_target.png", "width": 1376, "height": 768}
+{"index": 305, "id": "ab55ac6c-183b-475d-a951-c26224d89e40", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are an AI slide-generation engine. Convert the supplied slide JSON into a high-quality presentation image.\n\nFollow these constraints exactly:\n- The JSON is authoritative. Render only the contents of `blocks`; never add new words.\n- Embed all source images into the slide.\n- `global_style` defines the shared visual identity. Preserve its fonts, colors, background, and style language.\n- `visual_hierarchy` describes what should feel dominant, secondary, or subtle.\n- `special_instructions` may refine or override the default look; respect it when present.\n\nDesign with variety. Do not default to a title plus bullet list. Use asymmetry, columns, modular cards, large focal visuals, restrained annotations, or generous negative space when appropriate. The final slide must feel clean, readable, and presentation-ready.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Warm Artisan Baking\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#F6E7D2\\\",\\n      \\\"gradient\\\": \\\"radial gradient from #FDF6EC at center to #E2C8A3 at edges\\\",\\n      \\\"texture\\\": \\\"paper\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Playfair Display\\\",\\n      \\\"secondary\\\": \\\"Lato\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#8B4A2A\\\",\\n      \\\"secondary\\\": \\\"#C47A3A\\\",\\n      \\\"text\\\": \\\"#4A3323\\\",\\n      \\\"accent\\\": \\\"#D9A066\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"The slide should feel cozy and homely, echoing a rustic kitchen countertop. Use soft, warm beiges and browns that match the dough and chocolate filling in the image. Subtle paper texture in the background, with gentle shadows and rounded corners on cards to keep the focus on the baking process. Typography is elegant yet approachable: serif for the main title, clean sans-serif for supporting text. Overall atmosphere is warm, artisanal, and appetizing.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Warm, Homely, Artisanal\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 3,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-3\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Shaping a Chocolate Swirl Loaf\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#8B4A2A\\\",\\n            \\\"font_family\\\": \\\"Playfair Display\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, above the image with generous margin\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Large title, centered, with slightly increased letter-spacing to give an artisan cookbook feel.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/ab55ac6c-183b-475d-a951-c26224d89e40_input_0.png\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/ab55ac6c-183b-475d-a951-c26224d89e40_input_0.png\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Hands shaping a rolled pastry filled with rich chocolate spread on a countertop.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning about 75% of slide width and anchored in the middle\\\",\\n            \\\"image_style\\\": \\\"soft shadow with subtle warm-toned border\\\",\\n            \\\"container\\\": \\\"rounded rectangle frame in soft cream (#FDF6EC)\\\",\\n            \\\"special_instructions\\\": \\\"Apply a slight border-radius to the image corners (12px) and a warm, soft drop shadow to lift it off the background. The container frame should have thin border (#E0C19D) and minimal padding.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-3-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Roll, slice, and twist the dough to reveal rich layers of cocoa filling.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Lato\\\",\\n            \\\"color\\\": \\\"#4A3323\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image frame\\\",\\n            \\\"container\\\": \\\"pill-shaped tag with background #FDF6EC and thin border #D9A066\\\",\\n            \\\"container_opacity\\\": \\\"0.95\\\",\\n            \\\"special_instructions\\\": \\\"Small, centered text with medium weight, providing a single descriptive line. Add horizontal padding for the pill tag and keep it unobtrusive so the image remains dominant.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nUsing the provided slide JSON as the only source of truth, generate a clean, high-fidelity 16:9 presentation slide.", "input_images": ["images/multi_image/ab55ac6c-183b-475d-a951-c26224d89e40_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 306, "id": "a8b9ebc3-0893-4949-a326-8b2d47218f50", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "The slide presents a vibrant infographic in a paper-cutout art style, arranged into four tall vertical panels with a title banner in the upper left that now reads **\"FOUNDATIONS OF COMPUTER GRAPHICS SHADING.\"**  \n\nThe first panel, titled **\"SHADERS IN A NUTSHELL,\"** begins with a sub-image at the top showing a white cube in a blue and green corner with yellow rays. Beneath the image, the text explains shaders as **\"GPU-EXECUTED PROGRAMS\"** that **\"CONTROL HOW SURFACES ARE DRAWN ON SCREEN\"** and **\"COMPUTE COLOR, LIGHT RESPONSE, AND MATERIAL BEHAVIOR FOR EACH PIXEL.\"**  \n\nThe second panel, renamed **\"PRIMARY SHADER STAGES,\"** uses a winding ribbon graphic to separate two major stages: **\"VERTEX STAGE\"** (which **\"ADJUSTS VERTEX POSITIONS IN SPACE\"** and **\"PROCESSES PER-VERTEX DATA SUCH AS NORMALS AND UVS\"**) and **\"FRAGMENT / PIXEL STAGE\"** (which **\"DERIVES FINAL PIXEL COLOR\"** and **\"COMBINES TEXTURES, LIGHTING, AND MATERIAL PROPERTIES\"**).  \n\nThe third panel, now titled **\"BUILDING A SHADER,\"** organizes the workflow into three sections: **\"INPUT RESOURCES\"** (including **\"GEOMETRY BUFFERS,\" \"TEXTURE MAPS,\"** and **\"CONSTANT / UNIFORM PARAMETERS\"**), **\"SHADER PROGRAM\"** (described as **\"CODE WRITTEN IN GRAPHICS LANGUAGES SUCH AS GLSL, HLSL, OR Cg\"**), and **\"RENDERED RESULTS\"** (listing **\"PIXEL COLOR OUTPUT,\" \"DEPTH / Z-BUFFER DATA,\"** and **\"OPTIONAL AUXILIARY BUFFERS\"**).  \n\nThe rightmost panel, titled **\"POPULAR VISUAL TECHNIQUES,\"** highlights four groups of results achievable with shaders: **\"LIGHT & SHADOW MODELS\"** (including **\"LAMBERT, PHONG, PBR WORKFLOWS\"**), **\"SURFACE DETAILING\"** (with **\"ALBEDO, NORMAL, ROUGHNESS, AND METALLIC MAPPING\"**), **\"ENVIRONMENT & VOLUME EFFECTS\"** (such as **\"FOG, GOD RAYS, GLARE\"**), and **\"SCREEN-SPACE ENHANCEMENTS\"** (including **\"BLOOM, TONE MAPPING, ANTI-ALIASING, AND MOTION BLUR\"**).  \n\nAlong the bottom edge, a banner labeled **\"WHERE SHADERS ARE USED\"** points toward icons that represent **\"INTERACTIVE GAMES,\" \"CINEMATIC RENDERING & VFX,\"** and **\"TECHNICAL & SCIENTIFIC RENDERING.\"**", "input_images": ["images/multi_image/a8b9ebc3-0893-4949-a326-8b2d47218f50_input_0.jpg"], "target_image": "images/multi_image/a8b9ebc3-0893-4949-a326-8b2d47218f50_target.png", "width": 1376, "height": 768}
+{"index": 307, "id": "f0bc0d2b-979d-44b9-a0c9-ac7519b3b0d1", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This infographic, set against a background of green leaves with daisy illustrations, is titled \"SUPPORTING HEALTHY THYROID FUNCTION IN 3 SIMPLE STEPS\" in bold, black text at the top. The slide is organized into three numbered columns. The first column, labeled \"1,\" features the sub-image showing a fish, milk, an egg, and seaweed in the bottom-left position; the text reads \"NOURISH WITH FOOD\" followed by \"ADD NUTRIENT-DENSE OPTIONS. Include seafood, dairy, eggs, and sea vegetables regularly as part of a balanced eating pattern.\" The second column, labeled \"2,\" incorporates the sub-image of a salt shaker and a salt box in the middle; the text states \"CHOOSE FORTIFIED SALT\" followed by \"USE ENRICHED TABLE SALT. Look for wording on the package that indicates the salt has been fortified for thyroid support.\" The third column, labeled \"3,\" places the sub-image of a prescription bottle and a nursing mother in the bottom-right corner; the text reads \"FOLLOW MEDICAL GUIDANCE\" followed by \"DISCUSS SUPPLEMENTS WITH A CLINICIAN. People planning pregnancy, already pregnant, or breastfeeding should talk with their healthcare provider about appropriate vitamin and mineral supplementation.\"\n\n**IMPORTANT: 3 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/f0bc0d2b-979d-44b9-a0c9-ac7519b3b0d1_input_0.jpg", "images/multi_image/f0bc0d2b-979d-44b9-a0c9-ac7519b3b0d1_input_1.jpg", "images/multi_image/f0bc0d2b-979d-44b9-a0c9-ac7519b3b0d1_input_2.jpg"], "target_image": "images/multi_image/f0bc0d2b-979d-44b9-a0c9-ac7519b3b0d1_target.png", "width": 1376, "height": 768}
+{"index": 308, "id": "b94d5a82-8043-49af-a297-042a4af53555", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, designed in a sleek casino-night aesthetic, features a bold header reading **\"CARD HANDLING & GAME FLOW\"** in white and red typography, with a sub-header below stating **\"VISUALIZING KEY MOMENTS AT THE TABLE.\"** The main body is divided into three vertical panels, each showcasing a concise focus area built around the provided sub image of a person dealing playing cards.\n\n**Left Panel – \"DEAL\"**  \n- Use the sub image to depict the initial card distribution.  \n- Minimal caption: **\"Fair, smooth, consistent.\"**\n\n**Center Panel – \"DECIDE\"**  \n- Reuse the sub image cropped tightly on the Ace and Queen for emphasis.  \n- Minimal caption: **\"Read the table. Choose your move.\"**\n\n**Right Panel – \"REVEAL\"**  \n- Present the sub image again, highlighting the spread of cards on the surface.  \n- Minimal caption: **\"Show outcomes. Close the round.\"**\n\nA slim footer bar at the bottom reads **\"FOCUSED ON THE MOMENT OF THE DEAL\"** with subtle card-suit icons (♠ ♥ ♦ ♣) integrated into the design.", "input_images": ["images/multi_image/b94d5a82-8043-49af-a297-042a4af53555_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 309, "id": "5b8985f0-130b-4046-b4f5-6505bd7fb990", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### CREATIVE RENDERING BRIEF\nAct as a top-tier instructional designer generating a single troubleshooting slide from structured JSON.\n\n### JSON SEMANTICS\n- `global_style`: the fixed visual language for the entire deck. Apply it uniformly without alteration.\n- `visual_hierarchy`: defines viewing sequence, focal emphasis, and overall feel.\n- `source_images`: images that must appear on the slide.\n- `blocks`: the exclusive set of elements that may be rendered.\n- `special_instructions`: binding art direction that overrides general guidance when specified.\n\n### NON-NEGOTIABLES\n- Do not invent or alter any text, statistics, labels, icons, logos, captions, or section titles.\n- Ensure all text fits inside the visible canvas; fine-tune line breaks, size, spacing, and weight for clarity.\n- Honor all specified colors and fonts exactly. When a block references a palette color, use that precise value.\n- When content is sparse, preserve white space; do not introduce decorative elements or copy that are not in the JSON.\n\n### COMPOSITION GUIDANCE\nDesign a clear, modern layout tailored to the density of information. Use modular cards or columns for text-heavy areas; give priority space to illustrations on visual-focused sections; balance copy and imagery on mixed layouts so that the hierarchy is obvious at a glance. Deliver a polished, presentation-ready 16:9 slide image.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Corporate Educational Infographic\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#5B9BD5\",\n      \"secondary\": \"#F2F2F2\",\n      \"text\": \"#000000\",\n      \"accent\": \"#FFC000\"\n    },\n    \"special_instructions\": \"The slide uses a bright white canvas with a tidy grid framework. Content is grouped in rounded rectangular cards outlined in light grey. Visual styling is flat and illustrative, with cartoon-like characters and simple icons clarifying common presentation issues. The overall atmosphere is professional yet friendly, with a slight playful feel thanks to the character artwork.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Professional, Educational, Informative\",\n    \"composition_style\": \"Grid-based, 2x2 Quadrant Layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 134,\n      \"blocks\": [\n        {\n          \"id\": \"title-134\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"PROJECTOR TROUBLESHOOTING\",\n          \"style\": {\n            \"color\": \"#5B9BD5\",\n            \"font_family\": \"Arial\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"banner\",\n            \"special_instructions\": \"Place the title inside a light blue rounded banner with a darker blue stroke. Text is uppercase, centered, and visually prominent.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_80_180_350_413.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_80_180_350_413.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a man looking at a 'NO SIGNAL' screen\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-left quadrant, left side\",\n            \"image_style\": \"flat illustration\",\n            \"container\": \"rounded panel\",\n            \"special_instructions\": \"Belongs to the 'NO SIGNAL' module. Show a confused man facing a blue display labeled 'NO SIGNAL' with a separated projector cable.\"\n          }\n        },\n        {\n          \"id\": \"body-134-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"NO SIGNAL\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left quadrant, top right of panel\",\n            \"container\": \"blue rounded rectangle\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Use white uppercase text on a solid blue pill-shaped header, paired with a subtle connection icon.\"\n          }\n        },\n        {\n          \"id\": \"body-134-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"LOOSE VIDEO CABLE\\nINCORRECT INPUT MODE\\nVERIFY DEVICE OUTPUT\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left quadrant, below header\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Three-line checklist of quick checks, uppercase, left-aligned with comfortable line spacing.\"\n          }\n        },\n        {\n          \"id\": \"figure-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_722_183_989_407.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_722_183_989_407.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a man reacting to a burnt-out projector bulb\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"top-right quadrant, left side\",\n            \"image_style\": \"flat illustration\",\n            \"container\": \"rounded panel\",\n            \"special_instructions\": \"Belongs to the 'DIM OR BLANK IMAGE' module. Show a shocked man facing a screen with an 'X' while the projector glows faint orange.\"\n          }\n        },\n        {\n          \"id\": \"body-134-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"DIM OR BLANK IMAGE\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-right quadrant, top right of panel\",\n            \"container\": \"orange rounded rectangle\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Section label in white on a vivid orange header bar, with a small warning arrow icon.\"\n          }\n        },\n        {\n          \"id\": \"body-134-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"AGING LAMP MODULE\\nOVERHEATING SHUTDOWN\\nREPLACE PROJECTOR LAMP\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-right quadrant, below header\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Concise list of probable causes and remedy, uppercase, left-aligned beneath the header.\"\n          }\n        },\n        {\n          \"id\": \"figure-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_75_480_381_721.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_75_480_381_721.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of two people reacting to a system update pop-up\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-left quadrant, left side\",\n            \"image_style\": \"flat illustration\",\n            \"container\": \"rounded panel\",\n            \"special_instructions\": \"Belongs to the 'ON-SCREEN INTERRUPTIONS' module. Depict two worried presenters facing a large 'SYSTEM UPDATE' dialog during a slideshow.\"\n          }\n        },\n        {\n          \"id\": \"body-134-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"ON-SCREEN INTERRUPTIONS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"bottom-left quadrant, top right of panel\",\n            \"container\": \"yellow rounded rectangle\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Black uppercase heading on a yellow rounded bar, paired with a subtle notification icon.\"\n          }\n        },\n        {\n          \"id\": \"body-134-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"AUTO UPDATES TRIGGERED\\nALERT WINDOWS VISIBLE\\nDISABLE OR POSTPONE PROMPTS\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-left quadrant, below header\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Three-line action list, uppercase, left-aligned with consistent spacing for easy scanning.\"\n          }\n        },\n        {\n          \"id\": \"figure-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_704_474_1018_728.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_704_474_1018_728.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a man reacting to a power loss with a black screen\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-right quadrant, left side\",\n            \"image_style\": \"flat illustration\",\n            \"container\": \"rounded panel\",\n            \"special_instructions\": \"Belongs to the 'NO POWER' module. Show a surprised presenter with a laptop, a black projection screen, and an unplugged power cord.\"\n          }\n        },\n        {\n          \"id\": \"body-134-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"NO POWER\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#FFFFFF\",\n            \"importance\": \"dominant\",\n            \"position\": \"bottom-right quadrant, top right of panel\",\n            \"container\": \"green rounded rectangle\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"White uppercase title on a solid green header bar, paired with a slim lightning icon.\"\n          }\n        },\n        {\n          \"id\": \"body-134-8\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"AC CABLE UNPLUGGED\\nTRIPPED ROOM CIRCUIT\\nSITE-WIDE POWER FAILURE\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right quadrant, below header\",\n            \"container\": null,\n            \"container_opacity\": null,\n            \"special_instructions\": \"Uppercase bullet-style lines, left-aligned, summarizing primary power checks.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nCreate the final 16:9 slide image from the JSON above, following all style and content rules exactly.", "input_images": ["images/multi_image/5b8985f0-130b-4046-b4f5-6505bd7fb990_input_0.jpg", "images/multi_image/5b8985f0-130b-4046-b4f5-6505bd7fb990_input_1.jpg", "images/multi_image/5b8985f0-130b-4046-b4f5-6505bd7fb990_input_2.jpg", "images/multi_image/5b8985f0-130b-4046-b4f5-6505bd7fb990_input_3.jpg"], "target_image": "images/multi_image/5b8985f0-130b-4046-b4f5-6505bd7fb990_target.png", "width": 1376, "height": 768}
+{"index": 310, "id": "164205eb-871e-44a1-9149-5d97a5981cc0", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### CREATIVE RENDERING BRIEF\nAct as a world-class presentation designer generating a single slide from structured JSON.\n\n### JSON SEMANTICS\n- `global_style`: the immutable design system for this slide deck. Apply it consistently and exactly.\n- `visual_hierarchy`: the intended reading order, emphasis, and atmosphere.\n- `source_images`: the images to embed in the slide.\n- `blocks`: the only renderable content elements.\n- `special_instructions`: explicit creative direction that takes priority when it exists.\n\n### NON-NEGOTIABLES\n- Do not hallucinate text, numbers, labels, logos, captions, or section names.\n- Keep all text within visible boundaries; adjust line breaks, scale, spacing, and weight to improve legibility.\n- Preserve referenced colors and fonts. When a block points to a palette color, use that exact color.\n- If the content is minimal, let the slide breathe. Do not fill empty space with invented graphics or copy.\n\n### COMPOSITION GUIDANCE\nCreate a visually distinct layout that suits the content density. Text-heavy slides can use structured columns or staggered cards; visual-heavy slides should give the visual element priority; mixed slides should balance imagery and text with clear hierarchy. Output a refined, high-fidelity slide image.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Efficient Transformer Flow\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#020617\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #0f172a 0%, #020617 55%, #000000 100%)\\\",\\n      \\\"texture\\\": \\\"noise\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Montserrat\\\",\\n      \\\"secondary\\\": \\\"Inter\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#38BDF8\\\",\\n      \\\"secondary\\\": \\\"#F97316\\\",\\n      \\\"text\\\": \\\"#E5E7EB\\\",\\n      \\\"accent\\\": \\\"#A855F7\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall atmosphere is technical, sleek, and modern, evoking an AI research presentation. Dark navy and near-black background with a soft radial glow behind the central diagram suggests depth and focus. Subtle noise texture adds a matte, screen-like feel. Accent colors pick up the orange nodes and blue layers in the image, with clean sans-serif typography for a professional machine-learning talk.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Technical, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 3,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-3\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Token-Wise Early Exiting in Transformers\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#E5E7EB\\\",\\n            \\\"font_family\\\": \\\"Montserrat\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with generous margin above the main image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Large, all-caps or small-caps style with increased letter-spacing; centered directly above the image, kept on a single line if possible.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/164205eb-871e-44a1-9149-5d97a5981cc0_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/164205eb-871e-44a1-9149-5d97a5981cc0_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Diagram of a transformer encoder with token-dependent early exit based on a hash table.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning about 75% of slide width and vertically dominating the middle\\\",\\n            \\\"image_style\\\": \\\"clean with subtle outer glow\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Slight drop shadow and a 12px border-radius to soften edges; ensure the diagram remains crisp and is the largest visual element on the slide.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-3-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Cache layers per token to skip redundant computation and accelerate inference.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Inter\\\",\\n            \\\"color\\\": \\\"#9CA3AF\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, below the image with comfortable spacing\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Single concise line, medium weight, centered alignment; smaller than title but clearly readable.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-3-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A soft blurred ring composed of segmented lines suggesting data flow, placed behind the central diagram.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette with cyan and violet\\\",\\n            \\\"position\\\": \\\"centered behind the image block, extending slightly beyond its bounds\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (around 0.25), wide radius, subtle glow using #38BDF8 fading into #A855F7 to reinforce the sense of motion and technology without overpowering the main image.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nGenerate a refined 16:9 slide image suitable for a professional deck, with clear hierarchy, accurate content, and no unnecessary clutter.", "input_images": ["images/multi_image/164205eb-871e-44a1-9149-5d97a5981cc0_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 311, "id": "6478db6c-baf0-4654-9fde-c0abb019345a", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "The slide features a warm-toned travel poster titled **\"ROMAN COLOSSEUM\"** in bold, classic serif lettering at the top, set against a muted parchment background with a thin decorative border. The layout is divided into three main columns: History, Architecture, and Experience, with a narrow bottom bar for brief visitor tips. In the center of the slide, a large sub image of the ancient amphitheater dominates the design, showing its curved stone arches under a deep blue sky.\n\nIn the left column, under **History**, short lines read:  \n- \"Opened: 80 AD\"  \n- \"Commissioned by: Emperor Vespasian\"  \n- \"Used for: Games & public events\"\n\nIn the middle column, labeled **Architecture**, the text notes:  \n- \"Elliptical amphitheater\"  \n- \"Concrete & travertine stone\"  \n- \"Tiers of arches & columns\"\n\nIn the right column, titled **Experience**, concise phrases appear:  \n- \"Walk the arena floor\"  \n- \"See underground passages\"  \n- \"Imagine the roaring crowds\"\n\nThe bottom bar, labeled **VISITOR NOTES**, contains three tiny icons with minimal captions: a clock icon with \"Arrive early,\" a ticket icon with \"Book in advance,\" and a camera icon with \"Respect preservation rules.\" The overall emphasis of the slide stays on the sub image, allowing the visual of the Colosseum to be the primary focus.", "input_images": ["images/multi_image/6478db6c-baf0-4654-9fde-c0abb019345a_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 312, "id": "6c2284e1-ebff-4bea-be69-c60e231c6cc2", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are an elite AI Slide Composition Engine. Your capabilities blend the imagination of a senior Visual Designer with the rigor of an Information Architect.\n\n**Your Goal:** Transform the incoming **Structured Content JSON** into a single, production-ready presentation slide image.\n\n### INPUT DATA INTERPRETATION\nA JSON object describing ONE slide will be provided. Interpret it as follows:\n1. **`global_style`**: Establishes the non-negotiable visual universe for the slide (background treatment, materials, palette).\n2. **`visual_hierarchy`**: Defines compositional priorities, rhythm, and mood.\n3. **`blocks`**: The concrete renderable elements (text, images, illustrations, shapes).\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n* **Core Directive:** Any block with `type: \"image\"` must render the referenced source image exactly as provided.\n* **Source Assets:** Images are attached with this prompt and must be the precise assets shown on the slide.\n* **Protection Rules:**\n  * Do **NOT** crop, warp, redraw, or otherwise alter internal pixels of the image.\n  * Do **NOT** substitute with AI-generated visuals or generic stock photography.\n  * Do **NOT** omit, fade out, or partially hide required images.\n* **Permitted Styling:** You may stylize the *frame* around the image (borders, shadows, glass effects, perspective) so long as the original pixels remain fully visible and legible.\n* **Multiple Assets:** When several images appear, they must all be present and arranged coherently (side‑by‑side, stacked, grid, or as dictated by `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n* **Source of Truth:** Only use text found in the JSON `content` fields.\n* **Constraints:**\n  * **No fabrication:** Do not add new titles, taglines, or annotations.\n  * **No filler:** If text is minimal, keep the slide spacious and refined instead of inventing extra copy.\n* **Typography Handling:** You may adjust sizing, line breaks, letter spacing, and weights for visual balance, but avoid dense, unreadable paragraphs.\n* **`role: \"script\"` Usage:** Any block labeled `script` is purely contextual; use it to infer tone and atmosphere only. Do **not** display script text unless explicitly ordered via `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n* **Priority Rule:** Whenever a `special_instructions` field exists (either in `global_style` or a specific block), treat it as the creative override.\n* **Behavior:**\n  * When conflicts arise, `special_instructions` wins over default assumptions.\n  * Use this field to drive environmental details such as lighting, textures, motion cues, or thematic metaphors.\n  * Example: If a default suggests “flat background” but `special_instructions` calls for “neon glow arcade vibe,” you render the neon arcade look.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n* **Background:** Honor the exact `background.color`, gradients, and patterns specified. The appearance must be identical across all slides that share this style.\n* **Typeface Rules:** Use `font_palette.primary` for headings/titles and `font_palette.secondary` for body, lists, and captions—no substitutions.\n* **Color Usage:** Any `color_ref` must map directly to the hex codes defined in `color_palette`; do not approximate or introduce new hues.\n* **Role-Based Sizing Consistency:**\n  * TITLE (`importance: \"dominant\"`) blocks share identical scale and anchor point across slides.\n  * CORE_POINTS/LISTS (`importance: \"subheading\"` or equivalent) use uniform typography settings.\n  * CAPTION elements retain a consistent, smaller style.\n* **Intent:** Independently rendered slides must still feel like a unified deck—same typographic rhythm, same structural grid, same chromatic logic.\n\n#### 4. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n* **Anti-Monotony Rule:** Avoid repeating a basic “headline + bullets” stack for every slide.\n* **Implementation:**\n  * **Text-centric slides:** Experiment with multi-column layouts, staggered blocks, or asymmetry instead of a single vertical column.\n  * **Image-centric slides:** Let imagery dominate; place any text as concise labels or accents.\n  * **Hybrid slides:** Interweave visuals and text using overlaps, layered cards, or split screens.\n  * **Whitespace as Design:** When elements are few, deliberately leave breathing room. Do not inflate fonts or add dummy content to fill space.\n* **Variation Techniques:**\n  * Shift alignment (left, right, centered, diagonal clusters).\n  * Alter shapes of containers (pills, tabs, cards) while keeping style rules intact.\n  * Redistribute visual weight: some slides title‑heavy, others illustration‑heavy.\n* **Outcome:** A dynamic sequence in which every slide feels distinct yet unmistakably part of the same visual system.\n\n### OUTPUT GENERATION\nRender a final 16:9 slide image that follows these directives and faithfully represents the JSON specification.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Memphis Pop Art\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": \"Decorative geometric shapes (triangles, squiggles, dots) scattered around the edges and corners in a Memphis design style.\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2B6CB0\",\n      \"secondary\": \"#F59E0B\",\n      \"text\": \"#000000\",\n      \"accent\": \"#E53E8E\"\n    },\n    \"special_instructions\": \"The slide embraces a bold, retro Memphis aesthetic with playful geometry. All major text blocks and containers use thick black outlines. The palette leans on saturated pinks, teals, oranges, and purples set against a bright white base. The layout is divided into three vertical columns, each featuring a colored header bar and a content panel combining an icon with bulleted text.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Playful, Energetic, Informative\",\n    \"composition_style\": \"Grid-based, Three-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 719,\n      \"blocks\": [\n        {\n          \"id\": \"title-719\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"GROWTH MINDSET\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"panel\",\n            \"special_instructions\": \"The title sits inside a thick blue rectangular frame with stylized 'L' shaped corner brackets. Text is large, uppercase, and perfectly centered.\"\n          }\n        },\n        {\n          \"id\": \"illustration-719-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative Memphis-style geometric shapes (triangles, squiggles, dots) scattered around the slide borders and corners.\",\n          \"style\": {\n            \"illustration_style\": \"decorative background shape\",\n            \"color_scheme\": \"Pink, Blue, Teal, Black\",\n            \"position\": \"edges and corners\",\n            \"special_instructions\": \"These motifs frame the central content without significantly overlapping cards or text.\"\n          }\n        },\n        {\n          \"id\": \"body-719-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"WHAT A GROWTH MINDSET IS\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"left column, top header\",\n            \"container\": \"rounded rectangle with solid green background (#48C774)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Uppercase, bold wording centered inside a green header card with a thick black outline.\"\n          }\n        },\n        {\n          \"id\": \"image-719-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_97_329_462_457.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_97_329_462_457.jpg\"\n              }\n            ],\n            \"source\": \"Icon of a head with gears and a lightbulb\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left column, below header\",\n            \"image_style\": \"flat icon with thick black outline\",\n            \"container\": \"pink background panel\",\n            \"special_instructions\": \"The icon is centered within a pink panel that matches the overall Memphis aesthetic.\"\n          }\n        },\n        {\n          \"id\": \"body-719-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Believe abilities can be developed\\nView challenges as opportunities\\nSee effort as a path to mastery\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left column, below image\",\n            \"container\": \"pink background panel (#F48FB1)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Use small hollow square bullets; text is left-aligned within the pink card.\"\n          }\n        },\n        {\n          \"id\": \"body-719-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"WHY A GROWTH MINDSET MATTERS\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"center column, top header\",\n            \"container\": \"rounded rectangle with solid orange background (#F59E0B)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Uppercase, bold wording centered in an orange header box with a heavy black outline.\"\n          }\n        },\n        {\n          \"id\": \"image-719-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_509_329_866_456.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_509_329_866_456.jpg\"\n              }\n            ],\n            \"source\": \"Icon of a person pointing up with stars and arrows\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center column, below header\",\n            \"image_style\": \"flat icon with thick black outline\",\n            \"container\": \"light blue background panel\",\n            \"special_instructions\": \"The image is centered inside a light blue panel, framed with a black stroke.\"\n          }\n        },\n        {\n          \"id\": \"body-719-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Persist longer through setbacks\\nImprove learning and performance\\nBoost motivation and resilience\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"center column, below image\",\n            \"container\": \"light blue background panel (#A0D2EB)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Bullets are small hollow squares; copy is left-aligned and comfortably spaced.\"\n          }\n        },\n        {\n          \"id\": \"body-719-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"HOW TO CULTIVATE A GROWTH MINDSET\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#000000\",\n            \"importance\": \"dominant\",\n            \"position\": \"right column, top header\",\n            \"container\": \"rounded rectangle with solid purple background (#9F7AEA)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Uppercase, bold text centered inside a purple header bar with a thick black outline.\"\n          }\n        },\n        {\n          \"id\": \"illustration-719-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Icon of a hand writing in a notebook with a pencil.\",\n          \"style\": {\n            \"illustration_style\": \"flat icon\",\n            \"color_scheme\": \"Yellow notebook, blue hand, black outline\",\n            \"position\": \"right column, below header\",\n            \"special_instructions\": \"The icon is centered in a teal-toned panel to match the column style.\"\n          }\n        },\n        {\n          \"id\": \"body-719-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Reframe ‘failure’ as useful feedback\\nUse “yet” when you feel stuck\\nCelebrate effort and small improvements\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, below image\",\n            \"container\": \"teal background panel (#319795)\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Left-aligned bullet list with small hollow square markers inside the teal card.\"\n          }\n        },\n        {\n          \"id\": \"caption-719-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"@BrightMindStudio\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"bottom-right corner\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"special_instructions\": \"Discrete, small caption functioning as a credit watermark.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```", "input_images": ["images/multi_image/6c2284e1-ebff-4bea-be69-c60e231c6cc2_input_0.jpg", "images/multi_image/6c2284e1-ebff-4bea-be69-c60e231c6cc2_input_1.jpg"], "target_image": "images/multi_image/6c2284e1-ebff-4bea-be69-c60e231c6cc2_target.png", "width": 1376, "height": 768}
+{"index": 313, "id": "3fd6e501-abc9-4898-98b0-3b7d097669ff", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Featuring a serene mountain-themed educational slide titled **\"EMBRACING HIGH ALTITUDES\"** in clean, modern font framed by simple line icons of peaks, compasses, and hiking boots. At the top center, a slim banner reads **\"NATURAL LANDSCAPES\"**. The main illustration is the input image, placed prominently as the primary visual, showing a vast range of rugged peaks and a narrow grassy ridge under a clear sky. To the right, a smaller sub image crops in on the central ridge and lone hiker, emphasizing scale and perspective.  \n\nBelow the visuals, three soft-green panels provide brief text:\n\n1. **SCENE AT A GLANCE**  \n   - High mountain ridge at sunrise or sunset  \n   - Contrast of rocky cliffs and alpine meadow  \n\n2. **KEY FEATURES**  \n   - Sharp relief and deep valleys  \n   - Grassy foreground with wildflowers  \n   - Lone figure for sense of scale  \n\n3. **TAKEAWAYS**  \n   - Mountains highlight earth’s dramatic geology  \n   - Light and elevation shape how we perceive space", "input_images": ["images/multi_image/3fd6e501-abc9-4898-98b0-3b7d097669ff_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 314, "id": "92872048-d0e5-477e-8618-6e2bd621f57e", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "This presentation slide features a clean, white background with a minimalist black-and-white design, organized into a header and a three-column grid layout. At the top center, the main title reads \"ANCIENT STONE CIRCLE\" in a large, bold, uppercase sans-serif font. Below the title, the content is divided into three distinct vertical sections, each enclosed within a thin black rectangular border to maintain a structured and balanced appearance.\n\nThe first column on the left is headed by the sub image, scaled down and framed by a fine black border. It shows the monumental stone circle standing on an open green plain beneath a dramatic, cloud-filled sky. Directly beneath this illustration is the section header \"1. SETTING\" in bold, uppercase text. Below this header, a compact list in smaller, non-bold, uppercase font appears: \"OPEN GRASSLAND\", \"DISTANT HORIZON\", and \"CHANGING SKY\". The items are evenly spaced, emphasizing the broad landscape surrounding the ancient stones.\n\nThe middle column begins again with the same sub image, this time slightly cropped to emphasize the central trilithons and their massive vertical and horizontal stones. This image is likewise enclosed within a thin black rectangular frame, ensuring visual consistency. Underneath this graphic is the section header \"2. STONE FORMATION\" in bold, uppercase letters. The concise list below highlights key visual aspects in non-bold, uppercase text: \"RING OF MEGALITHS\", \"VERTICAL & HORIZONTAL BLOCKS\", and \"WEATHERED SURFACES\". These short phrases direct attention to the core structural features of the stone circle while keeping text minimal.\n\nThe third column on the right also features the sub image, here focused more on the sky and the interplay of light over the stones and surrounding field. The image is bordered in black like the others, aligning all three columns visually. Below this image is the header \"3. ATMOSPHERE\" in bold, uppercase text. The accompanying list, again in non-bold, uppercase lettering, contains three brief descriptors: \"DRAMATIC CLOUDS\", \"SOFT EVENING LIGHT\", and \"TIMELESS SILHOUETTE\". These phrases underline the mood and lighting that the sub image conveys, while allowing the photograph itself to remain the dominant element.\n\nAcross the slide, the high-contrast black-on-white color scheme and consistent sans-serif typography keep the layout clean and unobtrusive, ensuring that the sub image—repeated and reframed within each column—remains the main focus of the presentation.", "input_images": ["images/multi_image/92872048-d0e5-477e-8618-6e2bd621f57e_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 315, "id": "55304ec9-8afe-4bd8-ae84-206f66c96c98", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### OBJECTIVE  \nConvert the provided Structured Content JSON into a polished, presentation-ready slide in 16:9 format.\n\n### TEXT SOURCE RULES  \nTreat the JSON as the single source of truth for all on-slide wording. Do not invent new phrases, headings, bullets, labels, or filler language. If the JSON offers minimal text, preserve that minimalism.\n\n### VISUAL THEME GUIDELINES  \nUse `global_style` as the master design language. Apply its background, palette, typefaces, and mood consistently across the slide. Respect any `special_instructions` for lighting, depth, textures, and layout focus to maintain the soft 3D educational feel.\n\n### CONTENT BLOCK RENDERING  \nDisplay each visible block according to its declared role and hierarchy. Use any `script` fields only as invisible guidance for mood or narrative, not as visible copy. Ensure all text is comfortably legible, with adequate spacing and line length.\n\n### IMAGE HANDLING & PRESERVATION  \nFor every block where `type` is `\"image\"`, you MUST place the referenced image on the slide, using the specified paths. Do not crop away essential content or replace these assets.\n\n### COMPOSITION & LAYOUT  \nCreate a deliberate, premium-feeling composition that matches the “Soft 3D Educational” style. Follow the declared grid and column structure, while allowing variation in emphasis (some regions more visual, others more textual). The slide should look ready to drop into a professional deck.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Soft 3D Educational\",\n    \"background\": {\n      \"color\": \"#F0F7FC\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#2C4A6B\",\n      \"secondary\": \"#E8F5E9\",\n      \"text\": \"#2C3E50\",\n      \"accent\": \"#FFD700\"\n    },\n    \"special_instructions\": \"The slide uses a calm pastel scheme with a light, uncluttered backdrop. 3D, softly shaded illustrations with rounded forms establish an approachable, instructional atmosphere. Distinct tinted panels organize the main content areas.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Educational, Friendly, Informative\",\n    \"composition_style\": \"Grid-based, Three-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 224,\n      \"blocks\": [\n        {\n          \"id\": \"title-224\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"CITIZENSHIP & COMMUNITY CARE\",\n          \"style\": {\n            \"color\": \"#2C4A6B\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, uppercase heading text aligned to the left.\"\n          }\n        },\n        {\n          \"id\": \"subtitle-224\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"HOW PEOPLE SUPPORT ONE ANOTHER\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left, below title\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Smaller uppercase subtitle text, left-aligned under the main title.\"\n          }\n        },\n        {\n          \"id\": \"illustration-224-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative light blue curved shape in the top-left corner, partially behind the title text.\",\n          \"style\": {\n            \"illustration_style\": \"decorative background shape\",\n            \"color_scheme\": \"light blue\",\n            \"position\": \"top-left corner\",\n            \"special_instructions\": \"Semi-transparent or soft edge, adds visual interest to the header area.\"\n          }\n        },\n        {\n          \"id\": \"illustration-224-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A decorative light blue curved shape in the top-right corner.\",\n          \"style\": {\n            \"illustration_style\": \"decorative background shape\",\n            \"color_scheme\": \"light blue\",\n            \"position\": \"top-right corner\",\n            \"special_instructions\": \"Balances the design with the left-side shape.\"\n          }\n        },\n        {\n          \"id\": \"illustration-224-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A small decorative leaf element in the bottom-left corner.\",\n          \"style\": {\n            \"illustration_style\": \"decorative element\",\n            \"color_scheme\": \"green\",\n            \"position\": \"bottom-left corner\",\n            \"special_instructions\": \"Small, subtle detail.\"\n          }\n        },\n        {\n          \"id\": \"illustration-224-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A small decorative leaf element in the top-right area.\",\n          \"style\": {\n            \"illustration_style\": \"decorative element\",\n            \"color_scheme\": \"green\",\n            \"position\": \"top-right area\",\n            \"special_instructions\": \"Small, subtle detail.\"\n          }\n        },\n        {\n          \"id\": \"illustration-224-5\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A small decorative icon at the bottom center.\",\n          \"style\": {\n            \"illustration_style\": \"decorative element\",\n            \"color_scheme\": \"blue and white\",\n            \"position\": \"bottom center\",\n            \"special_instructions\": \"Small, subtle detail.\"\n          }\n        },\n        {\n          \"id\": \"panel-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large rounded rectangular panel with a light blue background (#D6E4F5) containing the first section of content.\",\n          \"style\": {\n            \"illustration_style\": \"content container\",\n            \"color_scheme\": \"light blue\",\n            \"position\": \"left column, middle section\",\n            \"special_instructions\": \"Acts as a background for the first section's image and text.\"\n          }\n        },\n        {\n          \"id\": \"image-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_24_194_449_433.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_24_194_449_433.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of two people holding a heart with stars and a lightbulb\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"inside left panel, top half\",\n            \"image_style\": \"none\",\n            \"container\": \"light blue rounded panel\",\n            \"special_instructions\": \"3D isometric illustration style.\"\n          }\n        },\n        {\n          \"id\": \"section-title-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"CARING ACTIONS\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"inside left panel, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, uppercase section header.\"\n          }\n        },\n        {\n          \"id\": \"body-224-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"sharing kindness and support\\nhelping friends, neighbours, and classmates\\ncelebrating ideas that make life better\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"minor\",\n            \"position\": \"inside left panel, bottom half\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"panel-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large rounded rectangular panel with a light green background (#E8F5E9) containing the second section of content.\",\n          \"style\": {\n            \"illustration_style\": \"content container\",\n            \"color_scheme\": \"light green\",\n            \"position\": \"center column, middle section\",\n            \"special_instructions\": \"Acts as a background for the second section's image and text.\"\n          }\n        },\n        {\n          \"id\": \"image-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_519_223_688_434.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_519_223_688_434.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of hands holding coins with a plant growing\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"inside center panel, top half\",\n            \"image_style\": \"none\",\n            \"container\": \"light green rounded panel\",\n            \"special_instructions\": \"3D isometric illustration style.\"\n          }\n        },\n        {\n          \"id\": \"section-title-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"WAYS WE CONTRIBUTE\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"inside center panel, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, uppercase section header.\"\n          }\n        },\n        {\n          \"id\": \"body-224-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"SAVING: Putting money aside for shared goals\\nSHARING: Offering resources when others need help\\nGROWING TOGETHER: Turning small efforts into big change\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"minor\",\n            \"position\": \"inside center panel, bottom half\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points with bolded keywords, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"panel-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large rounded rectangular panel with a light yellow background (#FFF9C4) containing the third section of content.\",\n          \"style\": {\n            \"illustration_style\": \"content container\",\n            \"color_scheme\": \"light yellow\",\n            \"position\": \"right column, middle section\",\n            \"special_instructions\": \"Acts as a background for the third section's image and text.\"\n          }\n        },\n        {\n          \"id\": \"image-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_930_204_1166_433.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_930_204_1166_433.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a diverse group of people under a tree with houses\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"inside right panel, top half\",\n            \"image_style\": \"none\",\n            \"container\": \"light yellow rounded panel\",\n            \"special_instructions\": \"3D isometric illustration style.\"\n          }\n        },\n        {\n          \"id\": \"section-title-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"STRONGER NEIGHBOURHOODS\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"supporting\",\n            \"position\": \"inside right panel, below image\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, uppercase section header.\"\n          }\n        },\n        {\n          \"id\": \"body-224-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"building trust and safety\\nsupporting places where everyone belongs\\ncreating spaces for families to learn and play\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#2C3E50\",\n            \"importance\": \"minor\",\n            \"position\": \"inside right panel, bottom half\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points, left-aligned.\"\n          }\n        },\n        {\n          \"id\": \"image-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1205_515_1331_701.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1205_515_1331_701.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of a person holding a star\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"bottom-right corner, outside the panel\",\n            \"image_style\": \"none\",\n            \"container\": \"none\",\n            \"special_instructions\": \"3D isometric illustration style, acts as a decorative element.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```", "input_images": ["images/multi_image/55304ec9-8afe-4bd8-ae84-206f66c96c98_input_0.jpg", "images/multi_image/55304ec9-8afe-4bd8-ae84-206f66c96c98_input_1.jpg", "images/multi_image/55304ec9-8afe-4bd8-ae84-206f66c96c98_input_2.jpg", "images/multi_image/55304ec9-8afe-4bd8-ae84-206f66c96c98_input_3.jpg"], "target_image": "images/multi_image/55304ec9-8afe-4bd8-ae84-206f66c96c98_target.png", "width": 1376, "height": 768}
+{"index": 316, "id": "bbd5a6dc-3958-440d-9c67-36b278e0e450", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### OBJECTIVE\nConvert the provided Structured Content JSON into a single polished, high-impact presentation slide.\n\n### CONTENT RULES\nAll visible text must come directly from the JSON. Do not invent extra wording, labels, bullets, or filler phrases. If the JSON includes little text, keep the design visually airy and uncluttered.\n\n### STYLE RULES\nTreat `global_style` as the master design system. Respect its background, palette, typography, and mood. Apply any `special_instructions` carefully, including notes on atmosphere, lighting, texture, and layout focus.\n\n### BLOCK RENDERING\nDisplay each visible block according to its designated role and hierarchy. Use `script` blocks, when present, only as invisible guidance for tone and theme. Ensure all text blocks remain clear and comfortably spaced.\n\n### IMAGE HANDLING\nFor any block with `type: \"image\"`, you MUST place the referenced image in the slide output without alteration to its content.\n\n### LAYOUT GUIDELINES\nCompose a single 16:9 slide with intentional structure. Allow variety in emphasis: imagery may dominate, or text may lead, or elements may form a diagram, but the result must look cohesive, premium, and presentation-ready.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Playful Cartoon Educational\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Rounded MT Bold\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#29B6F6\",\n      \"secondary\": \"#FF6F61\",\n      \"text\": \"#000000\",\n      \"accent\": \"#FFD54F\"\n    },\n    \"special_instructions\": \"The slide features a bright, cheerful, and cartoonish aesthetic. The style is reminiscent of children's educational materials or comic books, utilizing bold outlines, vibrant flat colors, and friendly character illustrations. The overall mood is inviting and lighthearted.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Playful, Informative, Friendly\",\n    \"composition_style\": \"Split layout with dominant central illustration and informational sidebar\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 779,\n      \"blocks\": [\n        {\n          \"id\": \"title-779\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"EXPLORE PLANT POWER WITH DORAEMON!\",\n          \"style\": {\n            \"color\": \"#29B6F6\",\n            \"font_family\": \"Arial Rounded MT Bold\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-left\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"outline\"\n              ],\n              \"outline_color\": \"#FFFFFF\",\n              \"shadow_type\": \"drop\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Large, bubbly font with a white outline and subtle drop shadow to make it pop against the white background.\"\n          }\n        },\n        {\n          \"id\": \"body-779-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Gentle wellness made from herbs!\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"left side, inside speech bubble\",\n            \"container\": \"white speech bubble with black outline\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a speech bubble pointing to Doraemon.\"\n          }\n        },\n        {\n          \"id\": \"body-779-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Ahh, this steam is amazing!\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-right, inside speech bubble\",\n            \"container\": \"white speech bubble with black outline\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a speech bubble pointing to the character Nobita.\"\n          }\n        },\n        {\n          \"id\": \"body-779-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"Fun for the whole gang!\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"bottom-right, inside speech bubble\",\n            \"container\": \"white speech bubble with black outline\",\n            \"container_opacity\": \"1.0\",\n            \"special_instructions\": \"Text is inside a speech bubble pointing to a group of characters.\"\n          }\n        },\n        {\n          \"id\": \"body-779-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"CALM YOUR MIND\\nHelps ease daily stress\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"right sidebar, top item\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold header followed by smaller regular text. Aligned to the right of a circular icon.\"\n          }\n        },\n        {\n          \"id\": \"body-779-5\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"FRESHEN YOUR SKIN\\nOpens pores, washes away dullness\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"right sidebar, second item\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold header followed by smaller regular text. Aligned to the right of a circular icon.\"\n          }\n        },\n        {\n          \"id\": \"body-779-6\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"ENERGISE YOUR BODY\\nKeeps blood moving smoothly\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"right sidebar, third item\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold header followed by smaller regular text. Aligned to the right of a circular icon.\"\n          }\n        },\n        {\n          \"id\": \"body-779-7\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"CLEAR STUFFY FEELINGS\\nHelps you breathe more easily\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"right sidebar, fourth item\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold header followed by smaller regular text. Aligned to the right of a circular icon.\"\n          }\n        },\n        {\n          \"id\": \"caption-779-1\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"Tip: Ask a healthcare professional if you are unsure whether steams are right for you.\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"position\": \"bottom-right corner\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"special_instructions\": \"Small font size, serving as a disclaimer.\"\n          }\n        },\n        {\n          \"id\": \"illustration-779-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A cartoon illustration of Doraemon on the left side, smiling and gesturing with one hand raised. He is blue with a white face, red nose, and a yellow bell on a red collar.\",\n          \"style\": {\n            \"illustration_style\": \"cartoon character\",\n            \"color_scheme\": \"blue, white, red, yellow\",\n            \"position\": \"left side, vertically centered\",\n            \"special_instructions\": \"Character is the focal point of the left side.\"\n          }\n        },\n        {\n          \"id\": \"illustration-779-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A wooden sauna cabin in the background center, with a bamboo fence to the right. Steam clouds are rising from a pot in the foreground.\",\n          \"style\": {\n            \"illustration_style\": \"cartoon background\",\n            \"color_scheme\": \"brown, green, beige\",\n            \"position\": \"center background\",\n            \"special_instructions\": \"Sets the scene for the herbal steam activity.\"\n          }\n        },\n        {\n          \"id\": \"illustration-779-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A large brown clay pot on a small stove in the foreground center. The pot contains water and herbs. Labels point to specific ingredients: 'BASIL', 'LEMONGRASS', 'KAFFIR LIME'.\",\n          \"style\": {\n            \"illustration_style\": \"cartoon object with labels\",\n            \"color_scheme\": \"brown, green, blue\",\n            \"position\": \"center foreground\",\n            \"special_instructions\": \"The labels are connected by thin black lines to the specific herbs in the pot.\"\n          }\n        },\n        {\n          \"id\": \"illustration-779-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A cartoon character (Nobita) sitting on the right side, wearing a white towel on his head and another around his neck. He looks happy and relaxed.\",\n          \"style\": {\n            \"illustration_style\": \"cartoon character\",\n            \"color_scheme\": \"skin tone, white, blue\",\n            \"position\": \"right side, next to the sauna\",\n            \"special_instructions\": \"Character is interacting with the scene, expressing enjoyment.\"\n          }\n        },\n        {\n          \"id\": \"illustration-779-5\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"A group of three cartoon characters (Shizuka, Gian, Suneo) in the bottom right corner, smiling and looking happy.\",\n          \"style\": {\n            \"illustration_style\": \"cartoon characters\",\n            \"color_scheme\": \"various\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"Small group illustration reinforcing the 'Great for everyone' message.\"\n          }\n        },\n        {\n          \"id\": \"illustration-779-6\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Four circular icons in the right sidebar, each depicting a benefit: a person in water (relax), a person washing face (skin), a person running (circulation), and a nose (congestion).\",\n          \"style\": {\n            \"illustration_style\": \"flat icons\",\n            \"color_scheme\": \"blue background with colorful illustrations\",\n            \"position\": \"right sidebar, vertically stacked\",\n            \"special_instructions\": \"Icons correspond to the text blocks describing the benefits.\"\n          }\n        },\n        {\n          \"id\": \"illustration-779-7\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative elements including small white cloud shapes and green leaves scattered around the slide.\",\n          \"style\": {\n            \"illustration_style\": \"decorative shapes\",\n            \"color_scheme\": \"white, green\",\n            \"position\": \"scattered throughout\",\n            \"special_instructions\": \"Adds to the light and airy feel of the slide.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```", "input_images": ["images/multi_image/bbd5a6dc-3958-440d-9c67-36b278e0e450_input_0.jpg"], "target_image": "images/multi_image/bbd5a6dc-3958-440d-9c67-36b278e0e450_target.png", "width": 1376, "height": 768}
+{"index": 317, "id": "40134ce5-4c40-4b85-8dc1-b9c3235610f9", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### MISSION\nTranslate the input Structured Content JSON into a single, refined visual poster with a cinematic, narrative feel.\n\n### STRICT CONTENT CONTROL\nAll visible wording must originate from the JSON. Do not invent taglines, annotations, interface labels, or filler phrases. If the JSON contains little text, preserve that sense of openness and negative space.\n\n### STYLE CONTROL\nTreat `global_style` as the master art direction. Adhere to its background treatment, palette, type choices, and overall ambience. Apply any `special_instructions` precisely, especially those that reference atmosphere, material texture, or focal layout.\n\n### BLOCK HANDLING\nInterpret each block according to its semantic role and stacking order. Use `script` blocks only as off-screen guidance for mood and story, never as visible copy. Ensure text blocks have generous breathing room and are easy to read at poster scale.\n\n### IMAGE INTEGRITY & PRESERVATION\nFor every block with `type: \"image\"`, you MUST place the referenced image files into the composition exactly once each, without cropping away essential content.\n\n### LAYOUT EXPECTATION\nCompose a 16:9 frame with intentional structure. Some posters may be image-dominant, others text-forward or emblematic. The final piece should appear polished, cohesive, and suitable for high-end print or digital display.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"New Chinese Aesthetics\",\n    \"background\": {\n      \"color\": \"#F7F6F2\",\n      \"gradient\": null,\n      \"texture\": \"paper\",\n      \"pattern\": \"Faint, low-opacity bamboo leaf silhouettes scattered across the background\"\n    },\n    \"font_palette\": {\n      \"primary\": \"Songti (Mincho)\",\n      \"secondary\": \"Songti (Mincho)\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#7B1111\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#1A1A1A\",\n      \"accent\": \"#D4AF37\"\n    },\n    \"special_instructions\": \"The piece carries a refined 'New Chinese Aesthetics' direction. The background mimics fibrous rice paper with a soft bamboo ghost pattern. Use a two-column balance with a large circular ink-wash illustration on the left and carefully spaced typographic content on the right. Integrate traditional red seals, gold cloud filigree, and a slim black frame. The atmosphere is serene, introspective, and steeped in classical culture.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Elegant, Cultural, Contemplative\",\n    \"composition_style\": \"Two-column (Image Left, Text Right)\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 187,\n      \"blocks\": [\n        {\n          \"id\": \"title-187\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"靜SERENITY\",\n          \"style\": {\n            \"color\": \"#7B1111\",\n            \"font_family\": \"Songti (Mincho)\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"The title pairs a large Chinese character '靜' (Serenity) in a traditional serif form with the English word 'SERENITY' in a bold serif face. Both appear in deep crimson red.\"\n          }\n        },\n        {\n          \"id\": \"subtitle-187\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"CULTIVATING STILLNESS IN A RESTLESS WORLD\",\n          \"style\": {\n            \"font_family\": \"Songti (Mincho)\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center, below title\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Subtitle set in all caps, black serif, centered directly beneath the main title.\"\n          }\n        },\n        {\n          \"id\": \"figure-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_111_241_653_703.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_111_241_653_703.jpg\"\n              }\n            ],\n            \"source\": \"Ink-wash painting of a figure holding a glowing orb\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"left-center, vertically centered\",\n            \"image_style\": \"Circular crop, traditional Chinese ink-wash style (Shui-mo hua)\",\n            \"container\": \"none\",\n            \"special_instructions\": \"The image depicts a robed figure in a misty mountain landscape holding a glowing golden orb. The style is monochromatic black and white with the orb as the only color accent. It is framed by decorative gold cloud motifs.\"\n          }\n        },\n        {\n          \"id\": \"illustration-187-1\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative gold cloud motifs (祥云) framing the circular image\",\n          \"style\": {\n            \"illustration_style\": \"Line art\",\n            \"color_scheme\": \"Gold (#D4AF37) outlines\",\n            \"position\": \"Surrounding the circular image on the left\",\n            \"special_instructions\": \"Traditional Chinese cloud patterns drawn with thin gold lines, appearing in the corners around the main image.\"\n          }\n        },\n        {\n          \"id\": \"illustration-187-2\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Red seal stamps (Chop) next to each list item\",\n          \"style\": {\n            \"illustration_style\": \"Icon\",\n            \"color_scheme\": \"Red (#7B1111)\",\n            \"position\": \"Left of each list item text on the right side\",\n            \"special_instructions\": \"Small vertical red rectangular seals with white Chinese characters inside, acting as bullet points.\"\n          }\n        },\n        {\n          \"id\": \"body-187-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"1. QUIET MIND: Breathing, observing, and allowing thoughts to settle like silt in clear water.\",\n          \"style\": {\n            \"font_family\": \"Songti (Mincho)\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, top\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Heading phrase is in bold red (#7B1111), supporting sentence in black. Preceded by a red seal icon and a small ink brush icon.\"\n          }\n        },\n        {\n          \"id\": \"body-187-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"2. GENTLE RHYTHMS: Honouring natural cycles of work, rest, and reflection.\",\n          \"style\": {\n            \"font_family\": \"Songti (Mincho)\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, below first item\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Heading phrase is in bold red (#7B1111), supporting sentence in black. Preceded by a red seal icon and a small cloud icon.\"\n          }\n        },\n        {\n          \"id\": \"body-187-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"3. INNER CLARITY: Listening beneath the noise to discern what truly matters.\",\n          \"style\": {\n            \"font_family\": \"Songti (Mincho)\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, middle\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Heading phrase is in bold red (#7B1111), supporting sentence in black. Preceded by a red seal icon and a small ink drop icon.\"\n          }\n        },\n        {\n          \"id\": \"body-187-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"4. DAILY PRACTICE: Small, consistent rituals that return the heart to calm.\",\n          \"style\": {\n            \"font_family\": \"Songti (Mincho)\",\n            \"color\": \"#1A1A1A\",\n            \"importance\": \"supporting\",\n            \"position\": \"right column, bottom\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Heading phrase is in bold red (#7B1111), supporting sentence in black. Preceded by a red seal icon and a small bamboo icon.\"\n          }\n        },\n        {\n          \"id\": \"footer-187\",\n          \"type\": \"text\",\n          \"role\": \"caption\",\n          \"content\": \"NEW CHINESE AESTHETICS | 16:9 POSTER\",\n          \"style\": {\n            \"font_family\": \"Songti (Mincho)\",\n            \"color\": \"#1A1A1A\",\n            \"position\": \"bottom-right\",\n            \"text_style\": {\n              \"font_weight\": \"normal\",\n              \"text_effect\": [],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"special_instructions\": \"Discrete footer text indicating the theme and canvas ratio.\"\n          }\n        },\n        {\n          \"id\": \"illustration-187-3\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Lotus flower line icon\",\n          \"style\": {\n            \"illustration_style\": \"Line art\",\n            \"color_scheme\": \"Black\",\n            \"position\": \"bottom-right corner\",\n            \"special_instructions\": \"A simple outline of a lotus flower, positioned to the left of the footer text.\"\n          }\n        },\n        {\n          \"id\": \"illustration-187-4\",\n          \"type\": \"illustration\",\n          \"role\": \"illustration\",\n          \"content\": \"Decorative border frame\",\n          \"style\": {\n            \"illustration_style\": \"Line art\",\n            \"color_scheme\": \"Black\",\n            \"position\": \"Top and bottom edges of the slide\",\n            \"special_instructions\": \"A thin black horizontal line with traditional geometric corner motifs at the top and bottom of the canvas.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nRender the structured slide JSON into a finalized 16:9 presentation image, preserving every character of text exactly and emphasizing a clear, tiered visual hierarchy.", "input_images": ["images/multi_image/40134ce5-4c40-4b85-8dc1-b9c3235610f9_input_0.jpg"], "target_image": "images/multi_image/40134ce5-4c40-4b85-8dc1-b9c3235610f9_target.png", "width": 1376, "height": 768}
+{"index": 318, "id": "a1654fbd-8d81-4e68-9b27-706c0ce1f9d6", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM\nYou are a cinematic presentation renderer with strong graphic design discipline.\n\n### OBJECTIVE\nGenerate a single polished slide image from the input JSON while preserving both semantic accuracy and deck-level style consistency.\n\n### RENDERING POLICY\n- Text fidelity is absolute: only render text from JSON `content` fields.\n- `global_style` controls the full visual language, including background, palette, fonts, texture, depth, lighting, and spacing.\n- `special_instructions` can introduce specific art direction such as atmosphere, materials, camera feel, or decorative treatment.\n- Maintain clear hierarchy: titles should lead, supporting text should be readable, captions should remain subtle, and visuals should not obscure essential text.\n- Use `source_images` to embed in the slide.\n\n### DESIGN DIRECTION\nChoose a layout that makes the slide memorable while staying coherent with the given style: bold hero composition, premium editorial layout, refined infographic, modular card layout, centered statement, or asymmetric grid. Keep the final output crisp, legible, balanced, and free of invented content.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Spring Blossom Serenity\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#FFF7FB\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top left, #FFE6F3 0%, #FFF7FB 55%, #FCEFFF 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Playfair Display\\\",\\n      \\\"secondary\\\": \\\"Source Sans Pro\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#D96BA8\\\",\\n      \\\"secondary\\\": \\\"#F3B3D4\\\",\\n      \\\"text\\\": \\\"#5A374A\\\",\\n      \\\"accent\\\": \\\"#F9D27D\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall atmosphere is soft, airy, and romantic, echoing a gentle spring morning. Use lots of white space and light pastel tones that match the cherry blossom pinks and warm golden leaves. Elements should feel delicate and slightly dreamy, with subtle shadows and rounded corners. No harsh borders; rely on translucency and soft gradients. The image is the clear hero, with text unobtrusively overlaid or floating above a translucent panel.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Elegant, Calm, Poetic\\\",\\n    \\\"composition_style\\\": \\\"Centered hero with overlaid title\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 3,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/a1654fbd-8d81-4e68-9b27-706c0ce1f9d6_input_0.png\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/a1654fbd-8d81-4e68-9b27-706c0ce1f9d6_input_0.png\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Close-up of pink cherry blossoms on a tree branch with a soft out-of-focus background.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center, spanning about 80% of slide width and 65% of height\\\",\\n            \\\"image_style\\\": \\\"soft shadow\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Use a wide 16:9 crop if needed, keeping blossoms as the focal point. Apply a very subtle drop shadow and gentle 12px border-radius to blend with the soft background. Ensure the image dominates the slide with minimal margins.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"title-3\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Whispers of Spring\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#5A374A\\\",\\n            \\\"font_family\\\": \\\"Playfair Display\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"overlaid near bottom-left of the image, avoiding the densest blossoms\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"banner\\\",\\n            \\\"special_instructions\\\": \\\"Title sits inside a translucent rounded rectangle (background #FFF7FB, opacity 0.78) that slightly overlaps the image. Left-aligned, large serif letters with generous letter spacing to feel airy.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-3-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Capturing the fleeting beauty of cherry blossoms in full bloom.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Source Sans Pro\\\",\\n            \\\"color\\\": \\\"#7B5468\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"directly below the title inside the same translucent banner\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Body text smaller and lighter in weight, left-aligned to the title with comfortable padding inside the banner; keep the line to a single sentence.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-3-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A soft, semi-transparent gradient circle that glows behind the title banner, using warm peach and pale gold hues.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"behind the title area at the lower-left portion of the image\\\",\\n            \\\"special_instructions\\\": \\\"Approximately 35% of slide width, low opacity (~0.35), slightly offset downward and to the left to create a subtle halo effect without distracting from the blossoms.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nTurn the structured slide JSON into a finished 16:9 presentation image, maintaining exact text fidelity and strong visual hierarchy.", "input_images": ["images/multi_image/a1654fbd-8d81-4e68-9b27-706c0ce1f9d6_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 319, "id": "a0bd3e78-2c7f-4787-bd39-ed6890cd1721", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This educational slide, titled \"REVENUE GROWTH: HOW FAST IS THE BUSINESS SCALING?\", utilizes Peanuts characters to explain financial concepts through a left-to-right narrative. The first sub-image, showing Charlie Brown, is positioned in the left column under the header \"STEP ONE: BASELINE REVENUE,\" accompanied by the text \"The amount of money a business is bringing in before a new growth push. For instance, from early customers and pilot sales. (Imagine it’s $500,000 per year).\" The second and third sub-images feature Snoopy and are stacked in the center column to illustrate the increase in sales, where the text reads \"GROWTH REVENUE = PRIOR REVENUE + NEW SALES ADDED\" and \"THE UPDATED RUN RATE: The company’s annualized revenue after landing new customers and upselling existing ones. (In this example: $500K Baseline + $300K New Sales) $800,000 Annual Revenue.\" The right side features Linus pointing to a bar chart labeled \"WHAT THIS SHOWS: MOMENTUM & TRACTION\" and Lucy at the bottom explaining the math with the text: \"Example breakdown: Existing customers ($500K), Newly acquired customers ($300K) for a total of $800K in revenue. Additional sales reduce dependence on any single buyer, but also require more support and delivery capacity. PERFORMANCE IS TRACKED PER PERIOD based on this updated revenue figure.\"", "input_images": ["images/multi_image/a0bd3e78-2c7f-4787-bd39-ed6890cd1721_input_0.jpg", "images/multi_image/a0bd3e78-2c7f-4787-bd39-ed6890cd1721_input_1.jpg", "images/multi_image/a0bd3e78-2c7f-4787-bd39-ed6890cd1721_input_2.jpg", "images/multi_image/a0bd3e78-2c7f-4787-bd39-ed6890cd1721_input_3.jpg"], "target_image": "images/multi_image/a0bd3e78-2c7f-4787-bd39-ed6890cd1721_target.png", "width": 1376, "height": 768}
+{"index": 320, "id": "3ac1b8ba-d6c9-4c34-bef7-aaee580902e0", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are the world's most advanced AI Presentation Rendering Engine. Your architecture combines the creativity of a top-tier Art Director with the precision of a Data Engineer.\n\n**Your Goal:** Convert the provided **Structured Content JSON** into a single, high-fidelity presentation slide image.\n\n### INPUT DATA INTERPRETATION\nYou will receive a JSON object representing ONE slide. You must parse it as follows:\n1.  **`global_style`**: Defines the immutable \"laws of physics\" for the slide (lighting, canvas material, color palette).\n2.  **`visual_hierarchy`**: Defines the compositional logic and mood.\n3.  **`blocks`**: The specific content elements (text, images, shapes) to be rendered.\n\n### CRITICAL EXECUTION PROTOCOLS (NON-NEGOTIABLE)\n\n#### 1. IMAGE INTEGRITY & PRESERVATION (HIGHEST PRIORITY)\n*   **The Rule:** When you encounter a block with `type: \"image\"`, you MUST embed the provided source images into your output.\n*   **Input Images:** Source images are provided as attachments along with this prompt. These are the ACTUAL images you must use.\n*   **Strict Constraint:** You must **STRICTLY PRESERVE** the internal pixel content of these source images.\n    *   **DO NOT** crop, distort, reimagine, or hallucinate the data inside the chart/table.\n    *   **DO NOT** replace the provided images with generic stock photos or AI-generated alternatives.\n    *   **DO NOT** ignore or omit the provided images - they MUST appear in the final slide.\n*   **Allowed Actions:** You MAY apply external styling to the *container* of the image (e.g., adding a glassmorphic border, a drop shadow, or a 3D perspective tilt) to make it fit the `theme_name`, provided the data remains legible.\n*   **Multiple Images:** If multiple images are provided, arrange them appropriately (side-by-side, stacked, or as specified in `special_instructions`).\n\n#### 2. TEXT ADAPTATION & BOUNDARIES\n*   **The Rule:** The JSON `content` field is your EXCLUSIVE source of text.\n*   **Strict Constraint:**\n    *   **NO Hallucinations:** Do not invent new text, headings, or summaries to fill space.\n    *   **NO Fluff:** If the content is sparse, embrace the negative space (whitespace) for a clean, high-end look.\n    *   **Formatting:** You are authorized to adjust typography (kerning, line breaks, font weights) for visual impact, but strictly prevent \"walls of text.\"\n*   **`role: \"script\"` Handling:** The `script` block is for **CONTEXT ONLY**. It tells you *what the slide is about* to inform your mood/lighting choices. **NEVER** render the script paragraph visually on the slide unless explicitly instructed in `special_instructions`.\n\n#### 3. THE \"SPECIAL_INSTRUCTIONS\" OVERRIDE\n*   **The Rule:** The key `special_instructions` (found in `global_style` or individual `blocks`) is the **Creative Override**.\n*   **Action:** If this key exists and contains text, it takes precedence over default rendering rules.\n    *   *Example:* If `global_style` says \"Blue Background\" but `special_instructions` says \"Immerse in a deep underwater bioluminescent scene,\" you execute the bioluminescent scene.\n    *   Use this field to add texture, lighting effects (bokeh, ray-tracing), and stylistic flair.\n\n#### 4. STRICT GLOBAL_STYLE ADHERENCE (CONSISTENCY)\n*   **The Rule:** The `global_style` defines the EXACT visual parameters that MUST be consistent across all slides.\n*   **Background:** Use EXACTLY the `background_color` specified. The background appearance must be IDENTICAL on every slide.\n*   **Fonts:** Use EXACTLY the fonts specified in `font_palette.primary` for titles and `font_palette.secondary` for body text.\n*   **Colors:** Use EXACTLY the hex codes from `color_palette` when `color_ref` references them.\n*   **Consistent Sizing by Role:**\n    *   All TITLE blocks (`font_size_weight: \"Dominant\"`) must appear at the SAME size and position on every slide.\n    *   All CORE_POINTS/LIST blocks (`font_size_weight: \"Subheading\"`) must use the SAME font style and size on every slide.\n    *   All CAPTION blocks (`font_size_weight: \"Caption\"`) must use the SAME font style and size on every slide.\n*   **Purpose:** This ensures that when multiple slides are generated independently, they will look like they belong to the same presentation - same title size, same bullet point style, same caption style.\n\n#### 4. LAYOUT DIVERSITY & VISUAL VARIETY (IMPORTANT)\n*   **The Rule:** AVOID the monotonous \"title + bullet list\" pattern. Each slide should feel visually unique.\n*   **Execution:**\n    *   **Text-heavy slides:** Arrange multiple body text blocks in creative layouts (columns, asymmetric placement, staggered positions) rather than a vertical list.\n    *   **Visual-led slides:** When images, tables, or illustrations dominate, let them take center stage. Body text (if any) should be minimal and positioned as supporting elements.\n    *   **Mixed slides:** Balance text and visuals with dynamic compositions—side-by-side, overlapping containers, or layered arrangements.\n    *   **Embrace whitespace:** If content is sparse, use negative space intentionally for a clean, premium aesthetic. Do NOT artificially stretch or pad content.\n*   **Variety Techniques:**\n    *   Vary text block positions (left-aligned, right-aligned, centered, scattered)\n    *   Use different container shapes or groupings per slide\n    *   Adjust visual weight distribution (sometimes title dominant, sometimes visual dominant)\n*   **Purpose:** Create a presentation that feels dynamic and engaging, where each slide has its own compositional identity while still adhering to `global_style` consistency.\n\n### OUTPUT GENERATION\nGenerate the final image representation of the slide based on these rules.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Streaming Efficiency Comparison\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#0B1020\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at top, #1C2440 0%, #050712 70%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Montserrat\\\",\\n      \\\"secondary\\\": \\\"Roboto\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#4A90E2\\\",\\n      \\\"secondary\\\": \\\"#F5A623\\\",\\n      \\\"text\\\": \\\"#F5F7FA\\\",\\n      \\\"accent\\\": \\\"#E95F5C\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"Overall look is a clean, data-focused tech presentation. The dark indigo background with a subtle radial glow behind the chart creates a dashboard feel and lets the light chart panel stand out. Use crisp sans-serif typography, thin gridlines if any, and minimal distractions. Corners of panels are slightly rounded, with soft drop shadows for depth. Accent colors echo the bar colors found in the chart (blue, gray, red, peach, ivory) for a cohesive appearance.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Professional, Technical, Focused\\\",\\n    \\\"composition_style\\\": \\\"Centered hero\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 5,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"title-5\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Streaming LLM Memory & Latency Benchmark\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#F5F7FA\\\",\\n            \\\"font_family\\\": \\\"Montserrat\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-center, with generous margin above the chart\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"none\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Title spans a single line when possible, centered, with slightly increased letter spacing for a modern tech feel.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/3ac1b8ba-d6c9-4c34-bef7-aaee580902e0_input_0.jpg\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/3ac1b8ba-d6c9-4c34-bef7-aaee580902e0_input_0.jpg\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Bar charts comparing memory usage and first-token latency across different streaming LLM methods and frame counts.\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"center of the slide, spanning about 75% of the width and 55% of the height\\\",\\n            \\\"image_style\\\": \\\"clean chart presentation with a subtle drop shadow\\\",\\n            \\\"container\\\": \\\"rounded white rectangle panel\\\",\\n            \\\"special_instructions\\\": \\\"Place the image inside a white panel with 10–12px rounded corners and a soft shadow (#000000 at 30% opacity, blur radius medium). Maintain the chart’s aspect ratio and ensure it is the dominant element on the slide.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-5-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"HERMES and StreamingTOM sustain low TTFT while significantly reducing memory across longer frame windows.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Roboto\\\",\\n            \\\"color\\\": \\\"#D6D9E5\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"bottom-center, just below the image panel\\\",\\n            \\\"container\\\": null,\\n            \\\"container_opacity\\\": null,\\n            \\\"special_instructions\\\": \\\"Set in small, centered text as a single concise takeaway. Keep line length short and use regular weight.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-5-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"A soft, oversized circular gradient halo positioned behind the chart panel, fading from bright blue at the center to transparent at the edges.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"gradient shape\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette with primary blue (#4A90E2) fading into the dark background\\\",\\n            \\\"position\\\": \\\"centered under the image panel\\\",\\n            \\\"special_instructions\\\": \\\"Low opacity (around 30–40%), larger than the image panel to create a subtle glow that draws attention to the chart without overpowering it.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nRender this JSON as a single 16:9 slide image. Make it visually compelling, readable, and consistent with all provided styling constraints.", "input_images": ["images/multi_image/3ac1b8ba-d6c9-4c34-bef7-aaee580902e0_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 321, "id": "df5e11af-4775-4603-9f7e-afde65fd22d1", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide, titled \"EXPLORING OUTPUT AND EARNINGS DYNAMICS\" in large, blue, uppercase letters at the top, visualizes how financial outcomes evolve as production volume changes. The layout features a detailed graph on the left and explanatory text on the right. The graph's vertical axis is labeled \"FINANCIAL VALUE ($ / Units)\" and the horizontal axis is \"PRODUCTION QUANTITY (Q).\" The chart displays three curves: \"TOTAL REVENUE,\" \"TOTAL COSTS,\" and \"FIXED COSTS.\" The area where costs are higher than revenue is shaded pink and labeled \"LOSS,\" containing a point marked \"LOSS MINIMIZATION POINT.\" The intersection of the revenue and cost lines is identified as the \"BREAK-EVEN POINT,\" leading to a \"BREAK-EVEN VOLUME\" on the x-axis. Beyond this, a blue shaded area represents \"PROFIT,\" peaking at the \"Maximum Profit\" corresponding to the \"MAXIMUM PROFIT OUTPUT.\" To the right, the \"CORE METRICS\" section lists \"Total Revenue (TR): Selling Price × Quantity Produced,\" \"Total Costs (TC): Sum of Fixed and Variable Expenditures,\" and \"Net Profit: Total Revenue – Total Costs.\" Below this, the \"ECONOMIC INSIGHTS\" section details the \"Break-Even Position (TR = TC): No gain or loss; all costs covered,\" the \"Profit-Generating Range: Output levels above Break-Even where earnings increase but may later taper off as inefficiencies arise,\" and \"Peak Profit: Achieved where additional revenue from the last unit exactly matches the additional cost of producing it (marginal revenue = marginal cost).\" The slide concludes with the \"OBJECTIVE: Determining the most efficient production level to enhance long-term financial performance.\"", "input_images": ["images/multi_image/df5e11af-4775-4603-9f7e-afde65fd22d1_input_0.jpg"], "target_image": "images/multi_image/df5e11af-4775-4603-9f7e-afde65fd22d1_target.png", "width": 1376, "height": 768}
+{"index": 322, "id": "d819dcfc-106c-4d5c-9271-6337a945fe35", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "The slide is presented with a distinct, retro pixel-art aesthetic, framed by a light brown, wood-textured border that mimics a computer window or a game interface, complete with a minimize and close icon in the top-right corner. The background behind this frame features a soft, pixelated blue sky with white blocky clouds, while the bottom edge reveals a strip of green grass. Dominating the upper section of the wooden frame is the main title \"COMMUNITY CHILDHOOD PROGRAMS\" written in a large, bold, dark brown pixelated font that stands out against the lighter wood background. Below the title, the content is organized into three vertical rectangular panels, each separated by a thin vertical line.\n\nOn the far left is the first panel, which features a beige header containing the text \"1. CREATIVE PLAY ZONES\" in a bold, dark brown pixel font. Beneath this header, the first sub-image depicts an indoor playroom scene where children are interacting with colorful building blocks and reading books near a bookshelf. In the center is the second panel, labeled with a beige header reading \"2. HEALTHY MEALS & WELLNESS\" in the same bold, dark brown style. The corresponding sub-image below illustrates a dining area where a server provides food to children seated at a table, with floating icons of healthy foods like an apple, bread, and a roasted chicken above them. To the right is the third and final panel, topped with a beige header displaying \"3. EARLY EDUCATION HUB\" in bold, dark brown letters. The sub-image in this section shows a classroom environment with a teacher reading to a group of children sitting on colorful mats in front of a chalkboard. Centered at the very bottom of the wooden frame, just below the three panels, is a small, pixelated icon of the Earth globe.\n\n**IMPORTANT: 1 source image(s) are attached. You MUST embed ALL of them into the slide.**", "input_images": ["images/multi_image/d819dcfc-106c-4d5c-9271-6337a945fe35_input_0.jpg"], "target_image": "images/multi_image/d819dcfc-106c-4d5c-9271-6337a945fe35_target.png", "width": 1376, "height": 768}
+{"index": 323, "id": "fc1bd176-4d2e-4985-ae13-a71e4774c38b", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "You are an AI slide-generation engine. Convert the supplied slide JSON into a high-quality presentation image.\n\nFollow these constraints exactly:\n- The JSON is authoritative. Render only the contents of `blocks`; never add new words.\n- Embed all source images into the slide.\n- `global_style` defines the shared visual identity. Preserve its fonts, colors, background, and style language.\n- `visual_hierarchy` describes what should feel dominant, secondary, or subtle.\n- `special_instructions` may refine or override the default look; respect it when present.\n\nDesign with variety. Do not default to a title plus bullet list. Use asymmetry, columns, modular cards, large focal visuals, restrained annotations, or generous negative space when appropriate. The final slide must feel clean, readable, and presentation-ready.\n\n### SLIDE JSON:\n```json\n\"{\\n  \\\"global_style\\\": {\\n    \\\"theme_name\\\": \\\"Focused Karate Energy\\\",\\n    \\\"background\\\": {\\n      \\\"color\\\": \\\"#0B1320\\\",\\n      \\\"gradient\\\": \\\"radial-gradient(circle at center, #1C2538 0%, #050812 100%)\\\",\\n      \\\"texture\\\": \\\"none\\\",\\n      \\\"pattern\\\": null\\n    },\\n    \\\"font_palette\\\": {\\n      \\\"primary\\\": \\\"Oswald\\\",\\n      \\\"secondary\\\": \\\"Open Sans\\\"\\n    },\\n    \\\"color_palette\\\": {\\n      \\\"primary\\\": \\\"#FFC93C\\\",\\n      \\\"secondary\\\": \\\"#FFFFFF\\\",\\n      \\\"text\\\": \\\"#F5F5F5\\\",\\n      \\\"accent\\\": \\\"#FF6B3D\\\"\\n    },\\n    \\\"special_instructions\\\": \\\"The slide has a dark, cinematic martial-arts atmosphere. The background uses a deep navy-to-black radial gradient that subtly brightens behind the central figure, mimicking a spotlight on the athlete. Typography is bold and condensed for the title, paired with a clean sans-serif for supporting text. Accents in yellow and warm orange echo the belt color and add energy. Overall feel: focused, powerful, minimal, with generous negative space around the dominant image.\\\"\\n  },\\n  \\\"visual_hierarchy\\\": {\\n    \\\"mood_tone\\\": \\\"Dynamic, Confident, Minimalist\\\",\\n    \\\"composition_style\\\": \\\"Two-column with large left hero image\\\"\\n  },\\n  \\\"styled_slides\\\": [\\n    {\\n      \\\"slide_id\\\": 19,\\n      \\\"blocks\\\": [\\n        {\\n          \\\"id\\\": \\\"figure-1\\\",\\n          \\\"type\\\": \\\"image\\\",\\n          \\\"role\\\": \\\"visual_elements\\\",\\n          \\\"content\\\": {\\n            \\\"image_paths\\\": [\\n              {\\n                \\\"path\\\": \\\"images/multi_image/fc1bd176-4d2e-4985-ae13-a71e4774c38b_input_0.png\\\",\\n                \\\"relative_path\\\": \\\"images/multi_image/fc1bd176-4d2e-4985-ae13-a71e4774c38b_input_0.png\\\"\\n              }\\n            ],\\n            \\\"source\\\": \\\"Person practicing a karate punch while wearing a white gi and yellow belt\\\",\\n            \\\"caption\\\": null\\n          },\\n          \\\"style\\\": {\\n            \\\"position\\\": \\\"left-center, occupying roughly 65% of slide width and vertically centered\\\",\\n            \\\"image_style\\\": \\\"subtle soft shadow to lift the figure from the dark background\\\",\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"Keep the image large and dominant, with a slight 10px border-radius on corners and a faint outer glow (#FFC93C at 20% opacity) on the right edge to emphasize motion.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"title-19\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"title\\\",\\n          \\\"content\\\": \\\"Harness Your Strike\\\",\\n          \\\"style\\\": {\\n            \\\"color\\\": \\\"#FFC93C\\\",\\n            \\\"font_family\\\": \\\"Oswald\\\",\\n            \\\"importance\\\": \\\"dominant\\\",\\n            \\\"position\\\": \\\"top-right, aligned vertically with the upper half of the image\\\",\\n            \\\"text_style\\\": {\\n              \\\"font_weight\\\": \\\"bold\\\",\\n              \\\"text_effect\\\": [\\n                \\\"shadow\\\"\\n              ],\\n              \\\"outline_color\\\": null,\\n              \\\"shadow_type\\\": \\\"drop\\\"\\n            },\\n            \\\"container\\\": \\\"none\\\",\\n            \\\"special_instructions\\\": \\\"All caps or small caps, tight letter-spacing, right-aligned to echo the forward punch direction.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"body-19-1\\\",\\n          \\\"type\\\": \\\"text\\\",\\n          \\\"role\\\": \\\"body\\\",\\n          \\\"content\\\": \\\"Precision, power, and presence in every movement.\\\",\\n          \\\"style\\\": {\\n            \\\"font_family\\\": \\\"Open Sans\\\",\\n            \\\"color\\\": \\\"#F5F5F5\\\",\\n            \\\"importance\\\": \\\"supporting\\\",\\n            \\\"position\\\": \\\"mid-right, below the title with comfortable spacing, right-aligned\\\",\\n            \\\"container\\\": \\\"rounded rectangle with semi-transparent dark overlay\\\",\\n            \\\"container_opacity\\\": \\\"0.35\\\",\\n            \\\"special_instructions\\\": \\\"Use a small, unobtrusive text size so it supports but does not compete with the title or image.\\\"\\n          }\\n        },\\n        {\\n          \\\"id\\\": \\\"illustration-19-1\\\",\\n          \\\"type\\\": \\\"illustration\\\",\\n          \\\"role\\\": \\\"illustration\\\",\\n          \\\"content\\\": \\\"Three short diagonal motion lines extending from the punching fist toward the right edge of the slide, suggesting speed.\\\",\\n          \\\"style\\\": {\\n            \\\"illustration_style\\\": \\\"minimal line art\\\",\\n            \\\"color_scheme\\\": \\\"matches accent palette\\\",\\n            \\\"position\\\": \\\"right edge of the slide, horizontally aligned with the extended fist\\\",\\n            \\\"special_instructions\\\": \\\"Use thin strokes in #FF6B3D with slight blur and 40% opacity; lines should be subtle and sit behind any text layers.\\\"\\n          }\\n        }\\n      ]\\n    }\\n  ]\\n}\"\n```\nRender the final output as one high-resolution 16:9 slide. Keep the slide free of hallucinated words, filler graphics, and unreadable text.", "input_images": ["images/multi_image/fc1bd176-4d2e-4985-ae13-a71e4774c38b_input_0.png"], "target_image": null, "width": 1376, "height": 768}
+{"index": 324, "id": "3c9b732d-937c-4174-96fd-0c4df3a27040", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Generate a slide titled **\"HEDGEHOG ATTENTION OVERVIEW\"** in large, bubbly teal lettering at the top center. The layout is divided into three vertical columns decorated with cute, cartoonish hedgehog and circuit-themed characters. The central focus of the slide is the provided sub image, placed prominently across the middle of all three columns, slightly overlapping them.\n\nThe first column, labeled **\"SOFTMAX ATTENTION\"**, briefly notes:\n- *Standard Q, K, V pipeline*\n- *Softmax over QKᵀ*\n- *MLP + Norm stack*\n\nThe middle column, labeled **\"TWO-STAGE DISTILLATION\"**, lists:\n- *Transfer from softmax to linear*\n- *Bridge between old & new blocks*\n- *Keeps behavior while changing form*\n\nThe third column, labeled **\"HEDGEHOG + MAMBA\"**, explains:\n- *φ(Q)φ(K)ᵀ linear attention*\n- *SSM-based mixing*\n- *Gated conv path for V*\n\nThe slide has a soft pastel background with abstract curved lines and subtle grid-like shapes to echo a neural-network feel. In the bottom right corner, the text **\"MODEL ARCHITECTURE SLIDE\"** appears in small caps.", "input_images": ["images/multi_image/3c9b732d-937c-4174-96fd-0c4df3a27040_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 325, "id": "aa508176-3770-44d1-8307-1f04f8c13a31", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "This presentation slide uses a modern, professional layout with a soft teal background, divided into clearly separated areas for charts and commentary. The slide is titled \"GLOBAL OMNI-CHANNEL PERFORMANCE OVERVIEW\" in a bold, dark sans-serif font across the top. Directly beneath the title, a compact metric strip displays icons and short figures: \"TOTAL GLOBAL COMMERCE VALUE: $[Figure] T\", \"DIGITAL SHARE OF SPEND: [Percentage]%\", and \"YEAR-ON-YEAR MOMENTUM: 📈 [Growth Description]\". \n\nThe central canvas is arranged into two vertical panels: on the left, the sub-image corresponding to the bar chart is embedded with the caption \"CHANNEL MIX BY MARKET: ONLINE VS STORE SALES (TOP ECONOMIES),\" highlighting how different countries allocate sales between digital and physical channels; on the right, the sub-image containing the radar chart is positioned beneath the heading \"STRUCTURAL DRIVERS OF ONLINE SALES BY CATEGORY,\" showing comparative scores across dimensions such as \"Digital Adoption\" and \"Consumer Trust.\" \n\nAt the bottom, two concise bullet lists synthesize the analysis: on the left, bullets read \"ONLINE CHANNELS CAPTURE A GROWING SHARE OF SPENDING\" and \"ASIA-PACIFIC SHOWS STRONGEST DIGITAL SALES MIX\"; on the right, bullets state \"LATENT DEMAND EXISTS WHERE INFRASTRUCTURE IS UNDERDEVELOPED\" and \"PAYMENTS, LOGISTICS, AND TRUST SHAPE CATEGORY PERFORMANCE.\" A slim footer in the lower right credits \"Source: Global Commerce Benchmark Data,\" while a centered notation at the very bottom shows the slide number \"22.4\".", "input_images": ["images/multi_image/aa508176-3770-44d1-8307-1f04f8c13a31_input_0.jpg", "images/multi_image/aa508176-3770-44d1-8307-1f04f8c13a31_input_1.jpg"], "target_image": "images/multi_image/aa508176-3770-44d1-8307-1f04f8c13a31_target.png", "width": 1376, "height": 768}
+{"index": 326, "id": "856837fa-d2cb-4116-a155-de2af247478d", "task": "multi_image", "source": "synthetic", "prompt_type": "short", "instruction": "Anchored by a teal gradient background and subtle geometric patterns, this slide presents a layout split between text on the left and a large illustrative image on the right. The prominent title \"Harborview Wind Project: Shaping a Greener Coast\" sits at the top in dark green text. Below the title, two white rounded rectangles contain the text \"This coastal wind initiative symbolizes the city’s long‑term commitment to clean energy.\" and \"Decades of innovation here have inspired broader regional investment in renewable power.\", each preceded by a circular icon featuring a line drawing of a leaf and a bar chart respectively. The right side features the main visual—a stylized illustration of a wind turbine overlooking a coastal cityscape with a golden plaque in the foreground—which corresponds to the input sub-image 1.", "input_images": ["images/multi_image/856837fa-d2cb-4116-a155-de2af247478d_input_0.jpg"], "target_image": "images/multi_image/856837fa-d2cb-4116-a155-de2af247478d_target.png", "width": 1376, "height": 768}
+{"index": 327, "id": "eb8efce8-ec17-45b6-88d7-5150f46fbaf0", "task": "multi_image", "source": "synthetic", "prompt_type": "long", "instruction": "Designing an educational slide titled **\"WHAT IS A CITY HALL?\" (Urban Spaces 101)**, the composition is arranged on a clean white background with a simple, modern aesthetic. At the very top center, the main title appears in bold, uppercase black sans‑serif font, followed directly underneath by a smaller subtitle, **\"CIVIC BUILDINGS & PUBLIC SPACES\"**, in a lighter black font. To the left of the subtitle sits a minimal dark‑green outline icon of a classical building, while to the right are two small geometric tree icons in matching green, echoing the greenery in the sub image.\n\nBeneath the heading area, the slide is structured into three evenly spaced vertical columns, each devoted to a short idea about what a city hall represents. The sub image is the central visual anchor of the slide and is placed prominently in the middle column, scaled large so it dominates the overall layout.\n\n---\n\n### Left Column – **\"1. PUBLIC PLACE\"**\n\n- At the top of the left column, the heading **\"1. PUBLIC PLACE\"** is written in bold black uppercase text, aligned center.\n- Directly underneath the heading, three short, minimal bullet lines explain the idea in small black font:\n  - \"Open to everyone.\"\n  - \"Center of local life.\"\n  - \"Meet, gather, participate.\"\n- No additional illustration appears here, leaving negative space so attention is drawn toward the sub image in the middle.\n\n---\n\n### Middle Column – Sub Image Focus – **\"CITY HALL\"**\n\n- The middle column is dedicated primarily to the sub image.\n- Centered within this column, the sub image shows a stylized, model‑like city hall building with a light cream façade, blue windows, and a flat blue roof. A small dome with a flag rises from the middle of the roof. Two tall, low‑poly trees stand on the right side of the building, and two simple stone obelisks frame the front area on a green base platform.\n- Below the sub image, a short caption in bold black uppercase letters reads **\"CITY HALL\"**.\n- Under the caption, in a smaller regular black font, a concise descriptive line is centered:\n  - \"Where a city is led and decisions are made.\"\n\n---\n\n### Right Column – **\"2. COMMUNITY & HISTORY\"** and **\"3. SHARED SPACE\"**\n\n- The upper part of the right column contains the heading **\"2. COMMUNITY & HISTORY\"** in bold black uppercase text, centered.\n- Under this heading, two short, centered lines of text in small black font:\n  - \"Symbol of the town.\"\n  - \"Records, ceremonies, stories.\"\n- A thin, light gray divider line separates this section from the lower part of the right column.\n- Beneath the divider, the second heading in this column reads **\"3. SHARED SPACE\"** in bold black uppercase.\n- Below it, three brief lines of text complete the slide:\n  - \"Public services inside.\"\n  - \"Open halls and rooms.\"\n  - \"Place for common decisions.\"\n\nThe overall layout keeps text minimal so the viewer’s attention is drawn primarily to the sub image of the city hall, using ample white space, clear typography, and a calm color palette that emphasizes civic order, openness, and shared urban identity.", "input_images": ["images/multi_image/eb8efce8-ec17-45b6-88d7-5150f46fbaf0_input_0.jpg"], "target_image": null, "width": 1376, "height": 768}
+{"index": 328, "id": "9b16787e-ccf5-45ce-abb6-2f8295357d27", "task": "multi_image", "source": "synthetic", "prompt_type": "style_json", "instruction": "### SYSTEM ROLE & OBJECTIVE\nYou are an elite AI Slide Composition Engine, fusing the sensibility of a seasoned Visual Designer with the rigor of an Information Architect.\n\n**Your Goal:** Transform the incoming **Structured Content JSON** into a single, visually refined 16:9 presentation slide image.\n\n### INPUT DATA INTERPRETATION\nA JSON object describing ONE slide will be provided. Interpret it as follows:\n1. **`global_style`**: Establishes the fundamental visual rules (lighting, surface, palette).\n2. **`visual_hierarchy`**: Defines layout logic, emphasis, and atmosphere.\n3. **`blocks`**: Individual elements (text, images, shapes) that must be arranged on the slide.\n\n### CORE EXECUTION DIRECTIVES (NON-NEGOTIABLE)\n\n#### 1. IMAGE FIDELITY & PRESERVATION (TOP PRIORITY)\n* **Mandate:** For any block with `type: \"image\"`, you MUST display the referenced source image in the final slide.\n* **Provided Assets:** Image files are attached to this prompt; these are the exact assets to be rendered.\n* **Strict Constraint:** You must **PRESERVE** the pixels inside each image.\n  * **DO NOT** crop, warp, redraw, or reinterpret any chart, illustration, or figure.\n  * **DO NOT** substitute with AI-generated images or stock photography.\n  * **DO NOT** omit any referenced image; every one must appear on the slide.\n* **Permitted Enhancements:** You MAY style the *frame* or container (e.g., borders, shadows, subtle tilts) as long as the image itself remains clear and unaltered.\n* **Multiple Images:** When several images are present, arrange them coherently according to layout hints or `special_instructions` (side‑by‑side, grid, stacked, etc.).\n\n#### 2. TEXT USAGE & LIMITS\n* **Mandate:** The JSON `content` fields are the ONLY textual sources you may render.\n* **Strict Constraint:**\n  * **No Inventions:** Do not add new titles, labels, or commentary beyond what appears in `content`.\n  * **No Padding:** Avoid filler phrases or decorative copy; empty areas should remain intentionally blank.\n  * **Formatting Flexibility:** You may adjust typography (line breaks, spacing, size, weight) for legibility and emphasis, while avoiding dense, unreadable blocks.\n* **`role: \"script\"` Handling:** If a block has `role: \"script\"`, treat it purely as background information for tone or style. **Never** place that script text onto the slide unless `special_instructions` explicitly require it.\n\n#### 3. THE `special_instructions` PRIORITY CHANNEL\n* **Mandate:** Any `special_instructions` field (in `global_style` or within a block) acts as a creative override.\n* **Action:** When present, it supersedes general defaults.\n  * Example: If a base rule suggests a flat background but `special_instructions` call for a parchment‑like texture, implement the parchment effect.\n  * Use this area to drive decisions on lighting, texture, micro‑layouts, and stylistic embellishments.\n\n#### 4. STRICT GLOBAL_STYLE CONSISTENCY\n* **Mandate:** `global_style` defines the canonical look; all elements must conform.\n* **Background:** Use the `background.color` exactly as given. The visual character of the background must remain stable.\n* **Fonts:** For headings, use `font_palette.primary`; for body and supporting text, use `font_palette.secondary`.\n* **Colors:** When `color_ref` is used, map it precisely to the specified `color_palette` hex values.\n* **Role-Based Sizing & Placement:**\n  * All TITLE blocks marked as `importance: \"dominant\"` must share consistent size and anchor position across slides.\n  * All CORE or LIST‑type body blocks (e.g., `importance: \"supporting\"` or designated subheads) must share font style and relative sizing.\n  * All CAPTION‑level elements must share a uniform caption style.\n* **Purpose:** This ensures slides produced independently still align as a single, cohesive deck.\n\n#### 5. LAYOUT VARIETY & COMPOSITION (IMPORTANT)\n* **Mandate:** Avoid repetitive “headline plus vertical list” layouts. Each slide should feel distinct yet related.\n* **Implementation:**\n  * **Text-centric slides:** Distribute text in columns or asymmetric groupings rather than stacking everything down the center.\n  * **Image-centric slides:** Let visuals dominate; reserve text for concise labels or summaries.\n  * **Hybrid slides:** Balance text and images with dynamic arrangements—overlapping layers, offset blocks, or mixed groupings.\n  * **Whitespace:** Treat empty space as a deliberate design element. Do not inflate fonts or add copy to fill gaps.\n* **Variation Techniques:**\n  * Shift text alignment (left, right, center) and position across the canvas.\n  * Use varied container shapes, panels, or subtle dividers.\n  * Modulate which element carries the most visual weight (sometimes title, sometimes imagery).\n* **Purpose:** Generate an engaging narrative flow where each slide has a unique structure but still respects the overarching `global_style`.\n\n### OUTPUT REQUIREMENTS\nRender a single, finished 16:9 slide image that adheres to these rules and fully reflects the JSON specification.\n\n### SLIDE JSON:\n```json\n{\n  \"global_style\": {\n    \"theme_name\": \"Educational Historical Profile\",\n    \"background\": {\n      \"color\": \"#FFFFFF\",\n      \"gradient\": null,\n      \"texture\": \"none\",\n      \"pattern\": null\n    },\n    \"font_palette\": {\n      \"primary\": \"Arial Black\",\n      \"secondary\": \"Arial\"\n    },\n    \"color_palette\": {\n      \"primary\": \"#000000\",\n      \"secondary\": \"#000000\",\n      \"text\": \"#000000\",\n      \"accent\": \"#000000\"\n    },\n    \"special_instructions\": \"The slide retains a bright white background with stark black typography and outlines. The overall look is that of a classroom handout turned into a polished infographic. A four-part grid organizes information around a single historical persona. The visuals are consistent, cartoon-style drawings of the same robed scholar figure, using soft browns, creams, and black linework for a unified, hand-drawn feel.\"\n  },\n  \"visual_hierarchy\": {\n    \"mood_tone\": \"Educational, Informative, Playful\",\n    \"composition_style\": \"Grid-based, Four-column layout\"\n  },\n  \"styled_slides\": [\n    {\n      \"slide_id\": 112,\n      \"blocks\": [\n        {\n          \"id\": \"title-112\",\n          \"type\": \"text\",\n          \"role\": \"title\",\n          \"content\": \"SU SHI (1037–1101) – LITERARY LUMINARY OF THE SONG\",\n          \"style\": {\n            \"color\": \"#000000\",\n            \"font_family\": \"Arial Black\",\n            \"importance\": \"dominant\",\n            \"position\": \"top-center\",\n            \"text_style\": {\n              \"font_weight\": \"bold\",\n              \"text_effect\": [\n                \"underline\"\n              ],\n              \"outline_color\": null,\n              \"shadow_type\": \"none\"\n            },\n            \"container\": \"none\",\n            \"special_instructions\": \"Strong, uppercase headline across the slide width, accented by a thick black underline.\"\n          }\n        },\n        {\n          \"id\": \"header-112-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"PUBLIC SERVICE & GOVERNANCE\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-left column\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, uppercase label for the first column.\"\n          }\n        },\n        {\n          \"id\": \"image-112-1\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_62_228_344_483.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_62_228_344_483.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of Su Shi writing at a desk\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-left column, below header\",\n            \"image_style\": \"cartoon illustration, flat colors\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Shows the robed scholar seated at a wooden table, brush in hand, composing text on a scroll.\"\n          }\n        },\n        {\n          \"id\": \"body-112-1\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Passed the imperial examinations with distinction\\n• Assisted in drafting key government documents\\n• Worked in multiple regional and central posts\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-left column\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Concise bullet list, left-aligned in the first column.\"\n          }\n        },\n        {\n          \"id\": \"header-112-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"POETRY & PROSE\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center-left column\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, uppercase label for the second column.\"\n          }\n        },\n        {\n          \"id\": \"image-112-2\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_398_228_664_488.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_398_228_664_488.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of Su Shi walking with clouds\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-center-left column, below header\",\n            \"image_style\": \"cartoon illustration, flat colors\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Shows the scholar strolling forward with swirling clouds and wind strokes around him.\"\n          }\n        },\n        {\n          \"id\": \"body-112-2\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Composed enduring ci lyrics and shi poems\\n• Wove nature, friendship, and politics into his writing\\n• Influenced later generations of East Asian literature\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-center-left column\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet points, left-aligned beneath the second illustration.\"\n          }\n        },\n        {\n          \"id\": \"header-112-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"BRUSHWORK & VISUAL ARTS\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-center-right column\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, uppercase label for the third column.\"\n          }\n        },\n        {\n          \"id\": \"image-112-3\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_719_229_1006_487.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_719_229_1006_487.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of Su Shi with a large brush and pine tree\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-center-right column, below header\",\n            \"image_style\": \"cartoon illustration, flat colors\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Shows the scholar holding an oversized calligraphy brush, with a pine tree and bold characters behind him.\"\n          }\n        },\n        {\n          \"id\": \"body-112-3\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Practiced distinctive calligraphy styles\\n• Advocated spontaneous, expressive strokes\\n• Inspired literati painting of rocks, trees, and distant vistas\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-center-right column\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Neat bullet list, aligned to the left in the third column.\"\n          }\n        },\n        {\n          \"id\": \"header-112-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"FOOD, LIFE & LEGEND\",\n          \"style\": {\n            \"font_family\": \"Arial Black\",\n            \"color\": \"#000000\",\n            \"importance\": \"supporting\",\n            \"position\": \"top-right column\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bold, uppercase label for the fourth column.\"\n          }\n        },\n        {\n          \"id\": \"image-112-4\",\n          \"type\": \"image\",\n          \"role\": \"visual_elements\",\n          \"content\": {\n            \"image_paths\": [\n              {\n                \"path\": \"imgs/img_in_image_box_1048_232_1315_488.jpg\",\n                \"relative_path\": \"imgs/img_in_image_box_1048_232_1315_488.jpg\"\n              }\n            ],\n            \"source\": \"Illustration of Su Shi eating pork\",\n            \"caption\": null\n          },\n          \"style\": {\n            \"position\": \"center-right column, below header\",\n            \"image_style\": \"cartoon illustration, flat colors\",\n            \"container\": \"none\",\n            \"special_instructions\": \"Shows the scholar smiling at a table set with a platter of braised pork and a teapot.\"\n          }\n        },\n        {\n          \"id\": \"body-112-4\",\n          \"type\": \"text\",\n          \"role\": \"body\",\n          \"content\": \"• Associated with rich braised pork dishes in later tales\\n• Became a beloved character in stories and opera\\n• Faced exile yet remained creatively active\",\n          \"style\": {\n            \"font_family\": \"Arial\",\n            \"color\": \"#000000\",\n            \"importance\": \"minor\",\n            \"position\": \"bottom-right column\",\n            \"container\": \"none\",\n            \"container_opacity\": null,\n            \"special_instructions\": \"Bullet list, left-aligned in the fourth column.\"\n          }\n        }\n      ]\n    }\n  ]\n}\n```\nNow render the final 16:9 slide image following these revised instructions, without adding any extra text beyond what is specified.", "input_images": ["images/multi_image/9b16787e-ccf5-45ce-abb6-2f8295357d27_input_0.jpg", "images/multi_image/9b16787e-ccf5-45ce-abb6-2f8295357d27_input_1.jpg", "images/multi_image/9b16787e-ccf5-45ce-abb6-2f8295357d27_input_2.jpg", "images/multi_image/9b16787e-ccf5-45ce-abb6-2f8295357d27_input_3.jpg"], "target_image": "images/multi_image/9b16787e-ccf5-45ce-abb6-2f8295357d27_target.png", "width": 1376, "height": 768}