Update README.md

README.md

@@ -9,4 +9,501 @@ base_model:
 pipeline_tag: image-text-to-image
 tags:
 - subject-driven
----
+---
+
+<p align="center">
+  <img src="assets/logo.png" alt="Scone" width="400"/>
+</p>
+<h3 align="center">
+    Scone: Bridging Composition and Distinction in Subject-Driven Image Generation </br>via Unified Understanding-Generation Modeling
+</h3>
+
+<p align="center"> 
+  <a href="https://arxiv.org/abs/2512.12675"><img alt="Build" src="https://img.shields.io/badge/arXiv%20paper-Scone-b31b1b.svg"></a>
+  <a href="https://github.com/Ryann-Ran/Scone"><img alt="Build" src="https://img.shields.io/github/stars/Ryann-Ran/Scone"></a> 
+  <a href="https://huggingface.co/Ryann829/Scone"><img src="https://img.shields.io/static/v1?label=%F0%9F%A4%97%20Hugging%20Face&message=Model&color=green"></a>
+  <a href="https://huggingface.co/datasets/Ryann829/Scone-S2I-57K"><img src="https://img.shields.io/static/v1?label=%F0%9F%A4%97%20Hugging%20Face&message=Data&color=yellow"></a>
+  <a href="https://huggingface.co/datasets/Ryann829/SconeEval"><img src="https://img.shields.io/static/v1?label=%F0%9F%A4%97%20Hugging%20Face&message=Benchmark&color=yellow"></a>
+
+
+
+
+
+# 🔧 Environment setup
+
+```bash
+git clone https://github.com/Ryann-Ran/Scone.git
+cd Scone
+conda create -n scone python=3.10 -y
+conda activate scone
+pip install -r requirements.txt
+pip install flash_attn==2.5.8 --no-build-isolation
+```
+
+
+# 🔍 Inference and Evaluation
+
+## Scone model preparation
+
+Download the Scone model checkpoint from [HuggingFace](https://huggingface.co/Ryann829/Scone):
+
+```bash
+# pip install -U huggingface_hub
+hf download Ryann829/Scone  --local-dir ./ckpts/Scone
+```
+
+## Single case inference
+
+Run the inference script:
+
+```bash
+bash scripts/inference_single_case.sh
+```
+
+**Example Output:** (Images sampled at 1024x1024 resolution with seed 1234, except for GPT-4o and Gemini-2.5-Flash-Image APIs)
+
+<table>
+  <tr>
+    <th style="width: 150px; text-align: center;">Ref. 1</th>
+    <th style="width: 150px; text-align: center;">Ref. 2</th>
+    <th style="width: 200px; text-align: center;">Instruction</th>
+    <th style="width: 150px; text-align: center;">Scone (Ours)</th>
+    <th style="width: 150px; text-align: center;">GPT-4o</th>
+    <th style="width: 150px; text-align: center;">Gemini-2.5-Flash-Image</th>
+    <th style="width: 150px; text-align: center;">UNO</th>
+    <th style="width: 150px; text-align: center;">Qwen-Image-Edit-2509</th>
+    <th style="width: 150px; text-align: center;">BAGEL</th>
+    <th style="width: 150px; text-align: center;">OmniGen2</th>
+    <th style="width: 150px; text-align: center;">Echo-4o</th>
+  </tr>
+  <tr>
+    <td style="width: 150px; text-align: center;"><img src="./example_images/image_object.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+    <td style="width: 150px; text-align: center;"><img src="./example_images/image_character.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+    <td style="width: 200px; text-align: center;">The man from image 2 holds the object which has a blue-and-red top in image 1 in a coffee shop.</td>
+    <td style="width: 150px; text-align: center;"><img src="./example_results/Scone.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+    <td style="width: 150px; text-align: center;"><img src="./example_results/GPT-4o.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+    <td style="width: 150px; text-align: center;"><img src="./example_results/Gemini-2.5-Flash-Image.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+    <td style="width: 150px; text-align: center;"><img src="./example_results/UNO.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+    <td style="width: 150px; text-align: center;"><img src="./example_results/Qwen-Image-Edit-2509.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+    <td style="width: 150px; text-align: center;"><img src="./example_results/BAGEL.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+    <td style="width: 150px; text-align: center;"><img src="./example_results/OmniGen2.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+    <td style="width: 150px; text-align: center;"><img src="./example_results/Echo-4o.png" style="width: 150px; height: auto; display: block; margin: 0 auto;" /></td>
+  </tr>
+</table>
+
+
+# 📊 Performance
+
+## OmniContext benchmark
+<table border="1" style="border-collapse: collapse; width: 100%;">
+    <thead>
+        <tr>
+            <th rowspan="2">Method</th>
+            <th colspan="2" style="text-align: center;">SINGLE ↑</th>
+            <th colspan="3" style="text-align: center;">MULTIPLE ↑</th>
+            <th colspan="3" style="text-align: center;">SCENE ↑</th>
+            <th rowspan="2" style="text-align: center;">Average ↑</th>
+        </tr>
+        <tr>
+            <th style="text-align: center;">Character</th>
+            <th style="text-align: center;">Object</th>
+            <th style="text-align: center;">Character</th>
+            <th style="text-align: center;">Object</th>
+            <th style="text-align: center;">Char. + Obj.</th>
+            <th style="text-align: center;">Character</th>
+            <th style="text-align: center;">Object</th>
+            <th style="text-align: center;">Char. + Obj.</th>
+        </tr>
+    </thead>
+    <tbody>
+        <tr>
+            <td colspan="10" style="background-color: #ffefe6; text-align: center; font-weight: bold; font-style: italic;">Closed-Source Model</td>
+        </tr>
+        <tr>
+            <td>Gemini-2.5-Flash-Image</td>
+            <td>8.79</td>
+            <td><strong>9.12</strong></td>
+            <td>8.27</td>
+            <td>8.60</td>
+            <td>7.71</td>
+            <td>7.63</td>
+            <td>7.65</td>
+            <td>6.81</td>
+            <td>8.07</td>
+        </tr>
+        <tr>
+            <td>GPT-4o<sup>*</sup></td>
+            <td><strong>8.96</strong></td>
+            <td>8.91</td>
+            <td><strong>8.90</strong></td>
+            <td><strong>8.95</strong></td>
+            <td><strong>8.81</strong></td>
+            <td><strong>8.92</strong></td>
+            <td><strong>8.40</strong></td>
+            <td><strong>8.44</strong></td>
+            <td><strong>8.78</strong></td>
+        </tr>
+        <tr>
+            <td colspan="10" style="background-color: #e0eef9; text-align: center; font-weight: bold; font-style: italic;">Generation Model</td>
+        </tr>
+        <tr>
+            <td>FLUX.1 Kontext [dev]</td>
+            <td>8.07</td>
+            <td>7.97</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+        </tr>
+        <tr>
+            <td>UNO</td>
+            <td>7.15</td>
+            <td>6.72</td>
+            <td>3.56</td>
+            <td>6.46</td>
+            <td>4.90</td>
+            <td>2.72</td>
+            <td>4.89</td>
+            <td>4.76</td>
+            <td>5.14</td>
+        </tr>
+        <tr>
+            <td>USO</td>
+            <td>8.03</td>
+            <td>7.55</td>
+            <td>3.32</td>
+            <td>6.10</td>
+            <td>4.56</td>
+            <td>2.77</td>
+            <td>5.38</td>
+            <td>5.09</td>
+            <td>5.35</td>
+        </tr>
+        <tr>
+            <td>UniWorld-V2</td>
+            <td>8.45</td>
+            <td><strong>8.44</strong></td>
+            <td>7.87</td>
+            <td>8.22</td>
+            <td><strong>7.95</strong></td>
+            <td><strong>5.36</strong></td>
+            <td>7.47</td>
+            <td><strong>6.98</strong></td>
+            <td>7.59</td>
+        </tr>
+        <tr>
+            <td>Qwen-Image-Edit-2509</td>
+            <td><strong>8.56</strong></td>
+            <td>8.41</td>
+            <td><strong>7.92</strong></td>
+            <td><strong>8.37</strong></td>
+            <td>7.79</td>
+            <td>5.23</td>
+            <td><strong>7.70</strong></td>
+            <td>6.86</td>
+            <td><strong>7.60</strong></td>
+        </tr>
+        <tr>
+            <td colspan="10" style="background-color: #E6E6FA; text-align: center; font-weight: bold; font-style: italic;">Unified Model</td>
+        </tr>
+        <tr>
+            <td>BAGEL</td>
+            <td>7.00</td>
+            <td>7.04</td>
+            <td>5.32</td>
+            <td>6.69</td>
+            <td>6.74</td>
+            <td>3.94</td>
+            <td>5.77</td>
+            <td>5.73</td>
+            <td>6.03</td>
+        </tr>
+        <tr>
+            <td>OmniGen2</td>
+            <td>8.17</td>
+            <td>7.63</td>
+            <td>7.26</td>
+            <td>7.03</td>
+            <td>7.56</td>
+            <td>7.02</td>
+            <td>6.90</td>
+            <td>6.64</td>
+            <td>7.28</td>
+        </tr>
+        <tr>
+            <td>Echo-4o</td>
+            <td><strong>8.34</strong></td>
+            <td>8.27</td>
+            <td>8.13</td>
+            <td><strong>8.14</strong></td>
+            <td>8.11</td>
+            <td><strong>7.07</strong></td>
+            <td>7.73</td>
+            <td><strong>7.77</strong></td>
+            <td>7.95</td>
+        </tr>
+        <tr>
+            <td><strong>Score (Ours)</strong></td>
+            <td><strong>8.34</strong></td>
+            <td><strong>8.52</strong></td>
+            <td><strong>8.24</strong></td>
+            <td><strong>8.14</strong></td>
+            <td><strong>8.30</strong></td>
+            <td>7.06</td>
+            <td><strong>7.88</strong></td>
+            <td>7.63</td>
+            <td><strong>8.01</strong></td>
+        </tr>
+    </tbody>
+</table>
+
+> - *: GPT-4o responded to 365~370 test cases out of the total 409 cases due to OpenAI safety restrictions.
+> - To mitigate randomness, we perform 3 rounds of sampling at 1024x1024 resolution, scoring 3 times per round, yielding 9 group results. The final score is the average of these results.
+
+
+## SconeEval benchmark
+<table border="1" style="border-collapse: collapse; width: 100%;">
+    <thead>
+        <tr>
+            <th rowspan="3">Method</th>
+            <th colspan="2">Composition ↑</th>
+            <th colspan="4">Distinction ↑</th>
+            <th colspan="4">Distinction & Composition ↑</th>
+            <th colspan="3">Average ↑</th>
+        </tr>
+        <tr>
+            <th>Single</th>
+            <th>Multi</th>
+            <th colspan="2">Cross</th>
+            <th colspan="2">Intra</th>
+            <th colspan="2">Cross</th>
+            <th colspan="2">Intra</th>
+            <th rowspan="2">COM</th>
+            <th rowspan="2">DIS</th>
+            <th rowspan="2">Overall</th>
+        </tr>
+        <tr>
+            <th>COM</th>
+            <th>COM</th>
+            <th>COM</th>
+            <th>DIS</th>
+            <th>COM</th>
+            <th>DIS</th>
+            <th>COM</th>
+            <th>DIS</th>
+            <th>COM</th>
+            <th>DIS</th>
+        </tr>
+    </thead>
+    <tbody>
+        <tr>
+            <td colspan="14" style="background-color: #ffefe6; text-align: center; font-weight: bold; font-style: italic;">Closed-Source Model</td>
+        </tr>
+        <tr>
+            <td>Gemini-2.5-Flash-Image</td>
+            <td>8.87</td>
+            <td>7.94</td>
+            <td>9.12</td>
+            <td><strong>9.15</strong></td>
+            <td>9.00</td>
+            <td>8.50</td>
+            <td>8.27</td>
+            <td><strong>8.87</strong></td>
+            <td>8.17</td>
+            <td>8.85</td>
+            <td>8.56</td>
+            <td>8.84</td>
+            <td>8.70</td>
+        </tr>
+        <tr>
+            <td>GPT-4o<sup>*</sup></td>
+            <td><strong>8.92</strong></td>
+            <td><strong>8.51</strong></td>
+            <td><strong>9.18</strong></td>
+            <td>8.55</td>
+            <td><strong>9.45</strong></td>
+            <td><strong>9.01</strong></td>
+            <td><strong>8.83</strong></td>
+            <td>8.49</td>
+            <td><strong>8.99</strong></td>
+            <td><strong>9.56</strong></td>
+            <td><strong>8.98</strong></td>
+            <td><strong>8.90</strong></td>
+            <td><strong>8.94</strong></td>
+        </tr>
+        <tr>
+            <td colspan="14" style="background-color: #e0eef9; text-align: center; font-weight: bold; font-style: italic;">Generation Model</td>
+        </tr>
+        <tr>
+            <td>FLUX.1 Kontext [dev]</td>
+            <td>7.92</td>
+            <td>-</td>
+            <td>7.93</td>
+            <td>8.45</td>
+            <td>6.20</td>
+            <td>6.11</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+            <td>-</td>
+        </tr>
+        <tr>
+            <td>USO</td>
+            <td>8.03</td>
+            <td>5.19</td>
+            <td>7.96</td>
+            <td>8.50</td>
+            <td>7.14</td>
+            <td>6.51</td>
+            <td>5.10</td>
+            <td>6.25</td>
+            <td>5.07</td>
+            <td>5.57</td>
+            <td>6.41</td>
+            <td>6.71</td>
+            <td>6.56</td>
+        </tr>
+        <tr>
+            <td>UNO</td>
+            <td>7.53</td>
+            <td>5.38</td>
+            <td>7.27</td>
+            <td>7.90</td>
+            <td>6.76</td>
+            <td>6.53</td>
+            <td>5.27</td>
+            <td>7.02</td>
+            <td>5.61</td>
+            <td>6.27</td>
+            <td>6.31</td>
+            <td>6.93</td>
+            <td>6.62</td>
+        </tr>
+        <tr>
+            <td>UniWorld-V2<br>(Edit-R1-Qwen-Image-Edit-2509)</td>
+            <td>8.41</td>
+            <td><strong>7.16</strong></td>
+            <td>8.63</td>
+            <td>8.24</td>
+            <td><strong>7.44</strong></td>
+            <td>6.77</td>
+            <td>7.52</td>
+            <td>8.03</td>
+            <td><strong>7.70</strong></td>
+            <td><strong>7.24</strong></td>
+            <td><strong>7.81</strong></td>
+            <td>7.57</td>
+            <td>7.69</td>
+        </tr>
+        <tr>
+            <td>Qwen-Image-Edit-2509</td>
+            <td><strong>8.54</strong></td>
+            <td>6.85</td>
+            <td><strong>8.85</strong></td>
+            <td><strong>8.57</strong></td>
+            <td>7.32</td>
+            <td><strong>6.86</strong></td>
+            <td><strong>7.53</strong></td>
+            <td><strong>8.13</strong></td>
+            <td>7.49</td>
+            <td>7.02</td>
+            <td>7.76</td>
+            <td><strong>7.65</strong></td>
+            <td><strong>7.70</strong></td>
+        </tr>
+        <tr>
+            <td colspan="14" style="background-color: #E6E6FA; text-align: center; font-weight: bold; font-style: italic;">Unified Model</td>
+        </tr>
+        <tr>
+            <td>BAGEL</td>
+            <td>7.14</td>
+            <td>5.55</td>
+            <td>7.49</td>
+            <td>7.95</td>
+            <td>6.93</td>
+            <td>6.21</td>
+            <td>6.44</td>
+            <td>7.38</td>
+            <td>6.87</td>
+            <td>7.27</td>
+            <td>6.74</td>
+            <td>7.20</td>
+            <td>6.97</td>
+        </tr>
+        <tr>
+            <td>OmniGen2</td>
+            <td>8.00</td>
+            <td>6.59</td>
+            <td>8.31</td>
+            <td>8.99</td>
+            <td>6.99</td>
+            <td>6.80</td>
+            <td>7.28</td>
+            <td>8.30</td>
+            <td>7.14</td>
+            <td>7.13</td>
+            <td>7.39</td>
+            <td>7.81</td>
+            <td>7.60</td>
+        </tr>
+        <tr>
+            <td>Echo-4o</td>
+            <td><strong>8.58</strong></td>
+            <td><strong>7.73</strong></td>
+            <td>8.36</td>
+            <td>8.33</td>
+            <td>7.74</td>
+            <td>7.18</td>
+            <td>7.87</td>
+            <td>8.72</td>
+            <td>8.01</td>
+            <td>8.33</td>
+            <td>8.05</td>
+            <td>8.14</td>
+            <td>8.09</td>
+        </tr>
+        <tr>
+            <td><strong>Scone (Ours)</strong></td>
+            <td>8.52</td>
+            <td>7.40</td>
+            <td><strong>8.98</strong></td>
+            <td><strong>9.73</strong></td>
+            <td><strong>7.97</strong></td>
+            <td><strong>7.74</strong></td>
+            <td><strong>8.20</strong></td>
+            <td><strong>9.25</strong></td>
+            <td><strong>8.21</strong></td>
+            <td><strong>8.44</strong></td>
+            <td><strong>8.21</strong></td>
+            <td><strong>8.79</strong></td>
+            <td><strong>8.50</strong></td>
+        </tr>
+    </tbody>
+</table>
+
+> - *: GPT-4o responded to 365~370 test cases out of the total 409 cases due to OpenAI safety restrictions.
+> - To mitigate randomness, we perform 3 rounds of sampling at 1024x1024 resolution, scoring 3 times per round, yielding 9 group results. The final score is the average of these results.
+
+
+
+#  🚰 Citation
+If you find Scone helpful, please consider giving the repo a star ⭐.
+
+If you find this project useful for your research, please consider citing our paper:
+```bibtex
+@misc{wang2025sconebridgingcompositiondistinction,
+      title={Scone: Bridging Composition and Distinction in Subject-Driven Image Generation via Unified Understanding-Generation Modeling}, 
+      author={Yuran Wang and Bohan Zeng and Chengzhuo Tong and Wenxuan Liu and Yang Shi and Xiaochen Ma and Hao Liang and Yuanxing Zhang and Wentao Zhang},
+      year={2025},
+      eprint={2512.12675},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2512.12675}, 
+}
+```