Hugging Face's logo

zai-org
/
GLM-Image

Text-to-Image
Diffusers
Safetensors
Chinese
English
GlmImagePipeline
Model card Files
xet
Community
13
ZHANGYUXUAN-zR commited on
Commit
1910a8d
·
verified ·
1 Parent(s): 16ff08b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +348 -48
README.md CHANGED
@@ -13,32 +13,44 @@ pipeline_tag: text-to-image
13
  <img src=https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/logo.svg width="40%"/>
14
  </div>
15
  <p align="center">
16
- 👋 Join our <a href="https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/wechat.jpeg" target="_blank">WeChat</a> and <a href="https://t.co/b6zGxJvzzS" target="_blank">Discord</a> community
17
  <br>
18
  📖 Check out GLM-Image's <a href="https://z.ai/blog/glm-image" target="_blank">Technical Blog</a>
19
  <br>
20
  📍 Use GLM-Image's <a href="https://docs.z.ai/guides/image/glm-image" target="_blank">API</a>
21
  </p>
22
 
23
- GLM-Image is an image generation model adopts a hybrid autoregressive + diffusion decoder architecture, effectively pushing the upper bound of visual fidelity and fine-grained details. In general image generation quality, it aligns with industry-standard LDM-based approaches, while demonstrating significant advantages in knowledge-intensive image generation scenarios.
24
 
25
- <div style="display: flex; justify-content: space-around;">
26
- <img src="https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/show_case.jpeg" width="45%"/>
27
- <img src="https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/show_case_t2i.jpe" width="45%"/>
28
- </div>
 
 
 
 
 
 
 
 
 
 
29
 
30
- Model architecture: a hybrid autoregressive + diffusion decoder design
31
 
32
- + Autoregressive generator: a 9B-parameter model initialized
33
- from [GLM-4-9B-0414](https://huggingface.co/zai-org/GLM-4-9B-0414), with an expanded vocabulary to incorporate visual tokens. The model first generates a compact encoding of approximately 256 tokens, then expands to 1K–4K tokens, corresponding to 1K–2K high-resolution image outputs.
 
 
 
34
  + Diffusion Decoder: a 7B-parameter decoder based on a single-stream DiT architecture for latent-space image decoding. It is equipped with a Glyph Encoder text module, significantly improving accurate text rendering within images.
35
 
36
  Post-training with decoupled reinforcement learning: the model introduces a fine-grained, modular feedback strategy using the GRPO algorithm, substantially enhancing both semantic understanding and visual detail quality.
37
 
38
  + Autoregressive module: provides low-frequency feedback signals focused on aesthetics and semantic alignment, improving instruction following and artistic expressiveness.
39
- + Decoder module: delivers high-frequency feedback targeting detail fidelity and text accuracy, resulting in highly realistic textures, lighting, and color reproduction, as well as more precise text rendering.
40
 
41
- GLM-Image supports both text-to-image and image-to-image generation within a single model
42
 
43
  + Text-to-image: generates high-detail images from textual descriptions, with particularly strong performance in information-dense scenarios.
44
  + Image-to-image: supports a wide range of tasks, including image editing, style transfer, multi-subject consistency, and identity-preserving generation for people and objects.
@@ -88,8 +100,8 @@ image = Image.open(image_path).convert("RGB")
88
  image = pipe(
89
  prompt=prompt,
90
  image=[image], # can input multiple images for multi-image-to-image generation such as [image, image1]
91
- height=33 * 32,
92
- width=32 * 32,
93
  num_inference_steps=30,
94
  guidance_scale=1.5,
95
  generator=torch.Generator(device="cuda").manual_seed(42),
@@ -98,45 +110,333 @@ image = pipe(
98
  image.save("output_i2i.png")
99
  ```
100
 
101
- + Since the AR model used in GLM-Image is configured with `do_sample=True` and a temperature of `0.95` by default, the generated images can vary significantly across runs. We do not recommend setting do_sample=False, as this may lead to incorrect or degenerate outputs from the AR model.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
102
 
103
- ### Prompt Enhancement for Generation
104
 
105
- We use GLM-4.7 to improve prompt, Please check our [github script](https://github.com/zai-org/GLM-Image/blob/main/examples/prompt_utils.py) for more details.
 
 
 
 
106
 
107
  ## Model Performance
108
 
109
  ### Text Rendering
110
 
111
- | **Model** | **Open Source** | **LongText-Bench-EN** | **LongText-Bench-ZH** | **CVTG-2K (Acc)** | **CVTG-2K (NED)** | **CVTG-2K (CLIPScore)** |
112
- |:------------------:|:---------------:|:---------------------:|:---------------------:|:-----------------:|:-----------------:|:-----------------------:|
113
- | Seedream 4.5 | ✗ | 0.989 | 0.987 | 0.8990 | 0.9483 | **0.8069** |
114
- | Seedream 4.0 | ✗ | 0.921 | 0.926 | 0.8451 | 0.9224 | 0.7975 |
115
- | Nano Banana 2.0 | ✗ | 0.981 | 0.949 | 0.7788 | 0.8754 | 0.7372 |
116
- | GPT Image 1 [High] | ✗ | 0.956 | 0.619 | 0.8569 | 0.9478 | 0.7982 |
117
- | Qwen-Image | ✓ | 0.943 | 0.946 | 0.8288 | 0.9116 | 0.8017 |
118
- | Qwen-Image-2512 | ✓ | 0.956 | 0.965 | 0.8604 | 0.9290 | 0.7819 |
119
- | Z-Image | ✓ | 0.935 | 0.936 | 0.8671 | 0.9367 | 0.7969 |
120
- | Z-Image-Turbo | ✓ | 0.917 | 0.926 | 0.8585 | 0.9281 | 0.8048 |
121
- | **GLM-Image** | ✓ | 0.952 | 0.979 | **0.9116** | **0.9557** | 0.7877 |
122
-
123
- ### Text-to-Image Benchmarks
124
-
125
- | **Model** | **Open Source** | **OneIG-Bench-EN** | **OneIG-Bench-ZH** | **TIIF-Bench short** | **TIIF-Bench long** | **DPG-Bench** |
126
- |--------------------|:---------------:|:------------------:|:------------------:|:--------------------:|:-------------------:|:-------------:|
127
- | Seedream 4.5 | ✗ | 0.576 | 0.551 | 90.49 | **88.52** | **88.63** |
128
- | Seedream 4.0 | ✗ | 0.576 | 0.553 | 90.45 | 88.08 | 88.54 |
129
- | Nano Banana 2.0 | ✗ | **0.578** | **0.567** | **91.00** | 88.26 | 87.16 |
130
- | GPT Image 1 [High] | ✗ | 0.533 | 0.474 | 89.15 | 88.29 | 85.15 |
131
- | DALL-E 3 | ✗ | - | - | 74.96 | 70.81 | 83.50 |
132
- | Qwen-Image | ✓ | 0.539 | 0.548 | 86.14 | 86.83 | 88.32 |
133
- | Qwen-Image-2512 | ✓ | 0.530 | 0.515 | 83.24 | 84.93 | 87.20 |
134
- | Z-Image | ✓ | 0.546 | 0.535 | 80.20 | 83.01 | 88.14 |
135
- | Z-Image-Turbo | ✓ | 0.528 | 0.507 | 77.73 | 80.05 | 84.86 |
136
- | FLUX.1 [Dev] | ✓ | 0.434 | - | 71.09 | 71.78 | 83.52 |
137
- | SD3 Medium | ✓ | - | - | 67.46 | 66.09 | 84.08 |
138
- | SD XL | ✓ | 0.316 | - | 54.96 | 42.13 | 74.65 |
139
- | BAGEL | ✓ | 0.361 | 0.370 | 71.50 | 71.70 | - |
140
- | Janus-Pro | ✓ | 0.267 | 0.240 | 66.50 | 65.01 | 84.19 |
141
- | Show-o2 | ✓ | 0.308 | - | 59.72 | 58.86 | - |
142
- | **GLM-Image** | ✓ | 0.528 | 0.511 | 81.01 | 81.02 | 84.78 |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
13
  <img src=https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/logo.svg width="40%"/>
14
  </div>
15
  <p align="center">
16
+ 👋 Join our <a href="https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/WECHAT.md" target="_blank">WeChat</a> and <a href="https://discord.gg/8KFjEec7" target="_blank">Discord</a> community
17
  <br>
18
  📖 Check out GLM-Image's <a href="https://z.ai/blog/glm-image" target="_blank">Technical Blog</a>
19
  <br>
20
  📍 Use GLM-Image's <a href="https://docs.z.ai/guides/image/glm-image" target="_blank">API</a>
21
  </p>
22
 
 
23
 
24
+ ## Case
25
+
26
+ ![show_case](https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/show_case.jpeg)
27
+
28
+ ### T2I with dense text and knowledge
29
+
30
+ ![show_case](https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/show_case_t2i.jpeg)
31
+
32
+ ### I2I
33
+
34
+ ![show_case](https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/show_case_i2i.jpeg)
35
+
36
+
37
+ ## Introduction
38
 
39
+ GLM-Image is an image generation model adopts a hybrid autoregressive + diffusion decoder architecture. In general image generation quality, GLM‑Image aligns with mainstream latent diffusion approaches, but it shows significant advantages in text-rendering and knowledge‑intensive generation scenarios. It performs especially well in tasks requiring precise semantic understanding and complex information expression, while maintaining strong capabilities in high‑fidelity and fine‑grained detail generation. In addition to text‑to‑image generation, GLM‑Image also supports a rich set of image‑to‑image tasks including image editing, style transfer, identity‑preserving generation, and multi‑subject consistency.
40
 
41
+ Model architecture: a hybrid autoregressive + diffusion decoder design.
42
+
43
+ ![architecture](https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/resources/architecture.jpeg)
44
+
45
+ + Autoregressive generator: a 9B-parameter model initialized from [GLM-4-9B-0414](https://huggingface.co/zai-org/GLM-4-9B-0414), with an expanded vocabulary to incorporate visual tokens. The model first generates a compact encoding of approximately 256 tokens, then expands to 1K–4K tokens, corresponding to 1K–2K high-resolution image outputs.
46
  + Diffusion Decoder: a 7B-parameter decoder based on a single-stream DiT architecture for latent-space image decoding. It is equipped with a Glyph Encoder text module, significantly improving accurate text rendering within images.
47
 
48
  Post-training with decoupled reinforcement learning: the model introduces a fine-grained, modular feedback strategy using the GRPO algorithm, substantially enhancing both semantic understanding and visual detail quality.
49
 
50
  + Autoregressive module: provides low-frequency feedback signals focused on aesthetics and semantic alignment, improving instruction following and artistic expressiveness.
51
+ + Decoder module: delivers high-frequency feedback targeting detail fidelity and text accuracy, resulting in highly realistic textures as well as more precise text rendering.
52
 
53
+ GLM-Image supports both text-to-image and image-to-image generation within a single model.
54
 
55
  + Text-to-image: generates high-detail images from textual descriptions, with particularly strong performance in information-dense scenarios.
56
  + Image-to-image: supports a wide range of tasks, including image editing, style transfer, multi-subject consistency, and identity-preserving generation for people and objects.
 
100
  image = pipe(
101
  prompt=prompt,
102
  image=[image], # can input multiple images for multi-image-to-image generation such as [image, image1]
103
+ height=33 * 32, # Must set height even it is same as input image
104
+ width=32 * 32, # Must set width even it is same as input image
105
  num_inference_steps=30,
106
  guidance_scale=1.5,
107
  generator=torch.Generator(device="cuda").manual_seed(42),
 
110
  image.save("output_i2i.png")
111
  ```
112
 
113
+ ### SGLang Pipeline
114
+
115
+ Install transformers and diffusers from source:
116
+
117
+ ```
118
+ pip install "sglang[diffusion] @ git+https://github.com/sgl-project/sglang.git#subdirectory=python"
119
+ pip install git+https://github.com/huggingface/transformers.git
120
+ pip install git+https://github.com/huggingface/diffusers.git
121
+ ```
122
+
123
+ + Text to Image Generation
124
+
125
+ ```
126
+ sglang serve --model-path zai-org/GLM-Image
127
+
128
+ curl http://localhost:30000/v1/images/generations \
129
+ -H "Content-Type: application/json" \
130
+ -d '{
131
+ "model": "zai-org/GLM-Image",
132
+ "prompt": "Doraemon is flying in the sky.",
133
+ "n": 1,
134
+ "response_format": "b64_json",
135
+ "size": "1024x1024"
136
+ }' | python3 -c "import sys, json, base64; open('output_t2i.png', 'wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"
137
+ ```
138
+
139
+ + Image to Image Generation
140
+
141
+ ```
142
+ sglang serve --model-path zai-org/GLM-Image
143
+
144
+ curl -s -X POST "http://localhost:30000/v1/images/edits" \
145
+ -F "model=zai-org/GLM-Image" \
146
+ -F "image=@cond.jpg" \
147
+ -F "prompt=Replace the background of the snow forest with an underground station featuring an automatic escalator." \
148
+ -F "response_format=b64_json" | python3 -c "import sys, json, base64; open('output_i2i.png', 'wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"
149
+ ```
150
 
151
+ ### Note
152
 
153
+ + We strongly recommend to use GLM-4.7 to enhance prompts for higher image quality, Please check [our github script](https://raw.githubusercontent.com/zai-org/GLM-Image/refs/heads/main/examples/prompt_utils.py) for more details.
154
+ + The AR model used in GLM‑Image is configured with `do_sample=True`, a temperature of `0.9`, and a topp of `0.75` by default. A higher temperature results in more diverse and rich outputs, but it can also lead to a certain decrease in output stability.
155
+ + The target image resolution must be divisible by 32. Otherwise, it will throw an error.
156
+ + Because the inference optimizations for this architecture are currently limited, the runtime cost is still relatively high. It requires either a single GPU with more than 80GB of memory, or a multi-GPU setup.
157
+ + vLLM-Omni and SGLang (with AR speedup) support is currently being integrated — stay tuned. For inference cost, you can check in our github.
158
 
159
  ## Model Performance
160
 
161
  ### Text Rendering
162
 
163
+ <div style="overflow-x: auto; margin-bottom: 16px;">
164
+ <table style="border-collapse: collapse; width: 100%;">
165
+ <thead>
166
+ <tr>
167
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa;" rowspan="2">Model</th>
168
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa;" rowspan="2">Open Source</th>
169
+ <th style="padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;" colspan="3">CVTG-2K</th>
170
+ <th style="padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;" colspan="3">LongText-Bench</th>
171
+ </tr>
172
+ <tr>
173
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">Word Accuracy</th>
174
+ <th style="padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">NED</th>
175
+ <th style="padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">CLIPScore</th>
176
+ <th style="padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">AVG</th>
177
+ <th style="padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">EN</th>
178
+ <th style="padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">ZH</th>
179
+ </tr>
180
+ </thead>
181
+ <tbody>
182
+ <tr>
183
+ <td style="padding: 8px; border: 1px solid #d0d7de;white-space:nowrap;">Seedream 4.5</td>
184
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✗</td>
185
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8990</td>
186
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.9483</td>
187
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>0.8069</strong></td>
188
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>0.988</strong></td>
189
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>0.989</strong></td>
190
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>0.987</strong></td>
191
+ </tr>
192
+ <tr>
193
+ <td style="padding: 8px; border: 1px solid #d0d7de;white-space:nowrap;">Seedream 4.0</td>
194
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✗</td>
195
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8451</td>
196
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.9224</td>
197
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.7975</td>
198
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.924</td>
199
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.921</td>
200
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.926</td>
201
+ </tr>
202
+ <tr>
203
+ <td style="padding: 8px; border: 1px solid #d0d7de;white-space:nowrap;">Nano Banana 2.0</td>
204
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✗</td>
205
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.7788</td>
206
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8754</td>
207
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.7372</td>
208
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.965</td>
209
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.981</td>
210
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.949</td>
211
+ </tr>
212
+ <tr>
213
+ <td style="padding: 8px; border: 1px solid #d0d7de;white-space:nowrap;">GPT Image 1 [High]</td>
214
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✗</td>
215
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8569</td>
216
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.9478</td>
217
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.7982</td>
218
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.788</td>
219
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.956</td>
220
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.619</td>
221
+ </tr>
222
+ <tr>
223
+ <td style="padding: 8px; border: 1px solid #d0d7de;white-space:nowrap;">Qwen-Image</td>
224
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
225
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8288</td>
226
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.9116</td>
227
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8017</td>
228
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.945</td>
229
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.943</td>
230
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.946</td>
231
+ </tr>
232
+ <tr>
233
+ <td style="padding: 8px; border: 1px solid #d0d7de;white-space:nowrap;">Qwen-Image-2512</td>
234
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
235
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8604</td>
236
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.9290</td>
237
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.7819</td>
238
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.961</td>
239
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.956</td>
240
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.965</td>
241
+ </tr>
242
+ <tr>
243
+ <td style="padding: 8px; border: 1px solid #d0d7de;white-space:nowrap;">Z-Image</td>
244
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
245
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8671</td>
246
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.9367</td>
247
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.7969</td>
248
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.936</td>
249
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.935</td>
250
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.936</td>
251
+ </tr>
252
+ <tr>
253
+ <td style="padding: 8px; border: 1px solid #d0d7de;white-space:nowrap;">Z-Image-Turbo</td>
254
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
255
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8585</td>
256
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.9281</td>
257
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.8048</td>
258
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.922</td>
259
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.917</td>
260
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.926</td>
261
+ </tr>
262
+ <tr>
263
+ <td style="padding: 8px; border: 1px solid #d0d7de;white-space:nowrap;"><strong>GLM-Image</strong></td>
264
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
265
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>0.9116</strong></td>
266
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>0.9557</strong></td>
267
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.7877</td>
268
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.966</td>
269
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.952</td>
270
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.979</td>
271
+ </tr>
272
+ </tbody>
273
+ </table>
274
+ </div>
275
+
276
+ ### Text-to-Image
277
+
278
+ <div style="overflow-x: auto; margin-bottom: 16px;">
279
+ <table style="border-collapse: collapse; width: 100%;">
280
+ <thead>
281
+ <tr>
282
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa;" rowspan="2">Model</th>
283
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa;" rowspan="2">Open Source</th>
284
+ <th style="padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;" colspan="2">OneIG-Bench</th>
285
+ <th style="padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;" colspan="2">TIIF-Bench</th>
286
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa;" rowspan="2">DPG-Bench</th>
287
+ </tr>
288
+ <tr>
289
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">EN</th>
290
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">ZH</th>
291
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">short</th>
292
+ <th style="white-space: nowrap; padding: 8px; border: 1px solid #d0d7de; background-color: #f6f8fa; text-align: center;">long</th>
293
+ </tr>
294
+ </thead>
295
+ <tbody>
296
+ <tr>
297
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">Seedream 4.5</td>
298
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✗</td>
299
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.576</td>
300
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.551</td>
301
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">90.49</td>
302
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>88.52</strong></td>
303
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>88.63</strong></td>
304
+ </tr>
305
+ <tr>
306
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">Seedream 4.0</td>
307
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✗</td>
308
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.576</td>
309
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.553</td>
310
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">90.45</td>
311
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">88.08</td>
312
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">88.54</td>
313
+ </tr>
314
+ <tr>
315
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">Nano Banana 2.0</td>
316
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✗</td>
317
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>0.578</strong></td>
318
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>0.567</strong></td>
319
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;"><strong>91.00</strong></td>
320
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">88.26</td>
321
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">87.16</td>
322
+ </tr>
323
+ <tr>
324
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">GPT Image 1 [High]</td>
325
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✗</td>
326
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.533</td>
327
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.474</td>
328
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">89.15</td>
329
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">88.29</td>
330
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">85.15</td>
331
+ </tr>
332
+ <tr>
333
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">DALL-E 3</td>
334
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✗</td>
335
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">-</td>
336
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">-</td>
337
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">74.96</td>
338
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">70.81</td>
339
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">83.50</td>
340
+ </tr>
341
+ <tr>
342
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">Qwen-Image</td>
343
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
344
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.539</td>
345
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.548</td>
346
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">86.14</td>
347
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">86.83</td>
348
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">88.32</td>
349
+ </tr>
350
+ <tr>
351
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">Qwen-Image-2512</td>
352
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
353
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.530</td>
354
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.515</td>
355
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">83.24</td>
356
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">84.93</td>
357
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">87.20</td>
358
+ </tr>
359
+ <tr>
360
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">Z-Image</td>
361
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
362
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.546</td>
363
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.535</td>
364
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">80.20</td>
365
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">83.01</td>
366
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">88.14</td>
367
+ </tr>
368
+ <tr>
369
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">Z-Image-Turbo</td>
370
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
371
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.528</td>
372
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.507</td>
373
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">77.73</td>
374
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">80.05</td>
375
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">84.86</td>
376
+ </tr>
377
+ <tr>
378
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">FLUX.1 [Dev]</td>
379
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
380
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.434</td>
381
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">-</td>
382
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">71.09</td>
383
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">71.78</td>
384
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">83.52</td>
385
+ </tr>
386
+ <tr>
387
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">SD3 Medium</td>
388
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
389
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">-</td>
390
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">-</td>
391
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">67.46</td>
392
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">66.09</td>
393
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">84.08</td>
394
+ </tr>
395
+ <tr>
396
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">SD XL</td>
397
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
398
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.316</td>
399
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">-</td>
400
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">54.96</td>
401
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">42.13</td>
402
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">74.65</td>
403
+ </tr>
404
+ <tr>
405
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">BAGEL</td>
406
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
407
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.361</td>
408
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.370</td>
409
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">71.50</td>
410
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">71.70</td>
411
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">-</td>
412
+ </tr>
413
+ <tr>
414
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">Janus-Pro</td>
415
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
416
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.267</td>
417
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.240</td>
418
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">66.50</td>
419
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">65.01</td>
420
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">84.19</td>
421
+ </tr>
422
+ <tr>
423
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;">Show-o2</td>
424
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
425
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.308</td>
426
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">-</td>
427
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">59.72</td>
428
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">58.86</td>
429
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">-</td>
430
+ </tr>
431
+ <tr>
432
+ <td style="padding: 8px; border: 1px solid #d0d7de; white-space:nowrap;font-weight:bold;">GLM-Image</td>
433
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">✓</td>
434
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.528</td>
435
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">0.511</td>
436
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">81.01</td>
437
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">81.02</td>
438
+ <td style="padding: 8px; border: 1px solid #d0d7de; text-align: center;">84.78</td>
439
+ </tr>
440
+ </tbody>
441
+ </table>
442
+ </div>