跳到正文
百度文心 ERNIE 官方博客·· 2026-10-22AI 評分68

百度發佈 ERNIE-Image 開放權重文生圖模型

Introducing ERNIE-Image

AI 導讀

百度發佈 ERNIE-Image,這是一款採用 8B 參數 DiT 的開放權重文生圖模型,並提供 ERNIE-Image-Turbo。ERNIE-Image 通常需 50 步推理,Turbo 經 DMD 和 RL 優化後只需 8 步。博客列出的四項基準結果中,ERNIE-Image 均位列前三,並在開放權重模型中排名第一。

正文

Image 1: ERNIEERNIE-Image

Overview Showcases Model Results Prompt Enhancer (PE)Conclusion

Get Started

Overview Showcases Model Results Prompt Enhancer (PE)Conclusion

Get Started

Introducing ERNIE‑Image

ERNIE-Image Team, Baidu

Image 2: ERNIE-Image Mosaic

1. Overview

ERNIE-Image is an open text-to-image model from the ERNIE-Image team at Baidu. Built on a single-stream Diffusion Transformer (DiT) with 8B parameters in a latent diffusion (LDM) framework, it ships with a lightweight Prompt Enhancer that expands brief inputs into richer, more structured prompts to better unlock the model's capabilities. With only 8B DiT parameters, ERNIE-Image achieves state-of-the-art performance among open weights text-to-image models — and it is built not just for visual appeal, but for controllability: accurate content depiction matters as much as aesthetics. In practice, it excels at complex instruction following, precise text rendering, and structured image generation — areas where many existing open weights models still fall short.

Key Features

  • •Competitive performance at compact scale: With only 8B DiT parameters, ERNIE-Image remains competitive with substantially larger models and achieves leading performance among open weights models on several challenging benchmarks.
  • •Precise text rendering: ERNIE-Image handles dense, long-form, and layout-sensitive text especially well, producing readable and faithful results in Chinese, English, and other languages.
  • •Robust instruction following: The model reliably handles complex prompts, multi-object relations, and knowledge-intensive descriptions, making it well suited for tasks that demand fine-grained control.
  • •Structured visual generation: ERNIE-Image is especially effective on images with clear layout or narrative structure — posters, manga/anime storyboards, multi-panel compositions, and cohesive multi-element visuals.
  • •Broad stylistic range: Beyond clean graphic design and illustration-style outputs, the model supports realistic photography and distinctive stylized aesthetics, including softer, more cinematic and film-like tones.
  • •Easy to deploy and adapt: Thanks to its compact size, ERNIE-Image runs on consumer-grade hardware (24G VRAM), bringing high-quality image generation within reach for research and production use. The moderate parameter count also makes fine-tuning and adaptation straightforward for researchers and developers.

Released Versions

Image 3: ERNIEERNIE-Image Our SFT model — stronger general-purpose quality and instruction fidelity, typically in 50 inference steps. 🤗 huggingface.co/Baidu/ERNIE-ImageImage 4: ERNIEERNIE-Image-Turbo Our Turbo model, optimized with DMD and RL for faster generation and stronger aesthetics in just 8 inference steps. 🤗 huggingface.co/Baidu/ERNIE-Image-Turbo

2. Showcases

The examples below highlight domains where ERNIE-Image stands out. Rather than focusing on generic image quality, we emphasize scenarios that demand text accuracy, structured composition, narrative organization, and distinctive visual style — settings where the gap between open weights models is most visible in practice. Click any image to view its full generation prompt. For more examples, visit our Prompt Gallery.

Image 5: Information Design sample 1

Image 6: Portrait & Lifestyle sample 2

Image 7: Information Design sample 3

Image 8: Information Design sample 4

Image 9: Concept & Imagination sample 4

Image 10: Posters & Character Worlds sample 4

Image 11: Narrative Illustration sample 5

Image 12: Posters & Character Worlds sample 5

Image 13: Portrait & Lifestyle sample 6

Image 14: Posters & Character Worlds sample 6

Image 15: Narrative Illustration sample 7

Image 16: Posters & Character Worlds sample 7

Image 17: Concept & Imagination sample 8

Image 18: Posters & Character Worlds sample 9

Image 19: Portrait & Lifestyle sample 10

Image 20: Information Design sample 11

Image 21: Narrative Illustration sample 11

Image 22: Information Design sample 12

Image 23: Concept & Imagination sample 12

Image 24: Information Design sample 13

Image 25: Concept & Imagination sample 13

Image 26: Portrait & Lifestyle sample 14

Image 27: Posters & Character Worlds sample 15

Image 28: Portrait & Lifestyle sample 17

Image 29: Information Design sample 19

Image 30: Information Design sample 21

Image 31: Information Design sample 23

Image 32: Portrait & Lifestyle sample 24

Image 33: Portrait & Lifestyle sample 25

Image 34: Information Design sample 27

Image 35: Information Design sample 1

Image 36: Portrait & Lifestyle sample 2

Image 37: Information Design sample 3

Image 38: Information Design sample 4

Image 39: Concept & Imagination sample 4

Image 40: Posters & Character Worlds sample 4

Image 41: Narrative Illustration sample 5

Image 42: Posters & Character Worlds sample 5

Image 43: Portrait & Lifestyle sample 6

Image 44: Posters & Character Worlds sample 6

Image 45: Narrative Illustration sample 7

Image 46: Posters & Character Worlds sample 7

Image 47: Concept & Imagination sample 8

Image 48: Posters & Character Worlds sample 9

Image 49: Portrait & Lifestyle sample 10

Image 50: Information Design sample 11

Image 51: Narrative Illustration sample 11

Image 52: Information Design sample 12

Image 53: Concept & Imagination sample 12

Image 54: Information Design sample 13

Image 55: Concept & Imagination sample 13

Image 56: Portrait & Lifestyle sample 14

Image 57: Posters & Character Worlds sample 15

Image 58: Portrait & Lifestyle sample 17

Image 59: Information Design sample 19

Image 60: Information Design sample 21

Image 61: Information Design sample 23

Image 62: Portrait & Lifestyle sample 24

Image 63: Portrait & Lifestyle sample 25

Image 64: Information Design sample 27

Image 65: Information Design sample 1

Image 66: Portrait & Lifestyle sample 2

Image 67: Information Design sample 3

Image 68: Information Design sample 4

Image 69: Concept & Imagination sample 4

Image 70: Posters & Character Worlds sample 4

Image 71: Narrative Illustration sample 5

Image 72: Posters & Character Worlds sample 5

Image 73: Portrait & Lifestyle sample 6

Image 74: Posters & Character Worlds sample 6

Image 75: Narrative Illustration sample 7

Image 76: Posters & Character Worlds sample 7

Image 77: Concept & Imagination sample 8

Image 78: Posters & Character Worlds sample 9

Image 79: Portrait & Lifestyle sample 10

Image 80: Information Design sample 11

Image 81: Narrative Illustration sample 11

Image 82: Information Design sample 12

Image 83: Concept & Imagination sample 12

Image 84: Information Design sample 13

Image 85: Concept & Imagination sample 13

Image 86: Portrait & Lifestyle sample 14

Image 87: Posters & Character Worlds sample 15

Image 88: Portrait & Lifestyle sample 17

Image 89: Information Design sample 19

Image 90: Information Design sample 21

Image 91: Information Design sample 23

Image 92: Portrait & Lifestyle sample 24

Image 93: Portrait & Lifestyle sample 25

Image 94: Information Design sample 27

Image 95: Portrait & Lifestyle sample 1

Image 96: Concept & Imagination sample 1

Image 97: Information Design sample 2

Image 98: Concept & Imagination sample 2

Image 99: Posters & Character Worlds sample 2

Image 100: Posters & Character Worlds sample 3

Image 101: Narrative Illustration sample 4

Image 102: Portrait & Lifestyle sample 5

Image 103: Information Design sample 6

Image 104: Concept & Imagination sample 6

Image 105: Information Design sample 7

Image 106: Information Design sample 8

Image 107: Portrait & Lifestyle sample 8

Image 108: Posters & Character Worlds sample 8

Image 109: Portrait & Lifestyle sample 9

Image 110: Concept & Imagination sample 9

Image 111: Concept & Imagination sample 10

Image 112: Portrait & Lifestyle sample 11

Image 113: Concept & Imagination sample 11

Image 114: Portrait & Lifestyle sample 12

Image 115: Posters & Character Worlds sample 12

Image 116: Information Design sample 14

Image 117: Posters & Character Worlds sample 14

Image 118: Portrait & Lifestyle sample 15

Image 119: Information Design sample 16

Image 120: Posters & Character Worlds sample 16

Image 121: Posters & Character Worlds sample 17

Image 122: Information Design sample 18

Image 123: Portrait & Lifestyle sample 20

Image 124: Portrait & Lifestyle sample 21

Image 125: Portrait & Lifestyle sample 22

Image 126: Information Design sample 24

Image 127: Information Design sample 26

Image 128: Portrait & Lifestyle sample 27

Image 129: Information Design sample 29

Image 130: Portrait & Lifestyle sample 1

Image 131: Concept & Imagination sample 1

Image 132: Information Design sample 2

Image 133: Concept & Imagination sample 2

Image 134: Posters & Character Worlds sample 2

Image 135: Posters & Character Worlds sample 3

Image 136: Narrative Illustration sample 4

Image 137: Portrait & Lifestyle sample 5

Image 138: Information Design sample 6

Image 139: Concept & Imagination sample 6

Image 140: Information Design sample 7

Image 141: Information Design sample 8

Image 142: Portrait & Lifestyle sample 8

Image 143: Posters & Character Worlds sample 8

Image 144: Portrait & Lifestyle sample 9

Image 145: Concept & Imagination sample 9

Image 146: Concept & Imagination sample 10

Image 147: Portrait & Lifestyle sample 11

Image 148: Concept & Imagination sample 11

Image 149: Portrait & Lifestyle sample 12

Image 150: Posters & Character Worlds sample 12

Image 151: Information Design sample 14

Image 152: Posters & Character Worlds sample 14

Image 153: Portrait & Lifestyle sample 15

Image 154: Information Design sample 16

Image 155: Posters & Character Worlds sample 16

Image 156: Posters & Character Worlds sample 17

Image 157: Information Design sample 18

Image 158: Portrait & Lifestyle sample 20

Image 159: Portrait & Lifestyle sample 21

Image 160: Portrait & Lifestyle sample 22

Image 161: Information Design sample 24

Image 162: Information Design sample 26

Image 163: Portrait & Lifestyle sample 27

Image 164: Information Design sample 29

Image 165: Portrait & Lifestyle sample 1

Image 166: Concept & Imagination sample 1

Image 167: Information Design sample 2

Image 168: Concept & Imagination sample 2

Image 169: Posters & Character Worlds sample 2

Image 170: Posters & Character Worlds sample 3

Image 171: Narrative Illustration sample 4

Image 172: Portrait & Lifestyle sample 5

Image 173: Information Design sample 6

Image 174: Concept & Imagination sample 6

Image 175: Information Design sample 7

Image 176: Information Design sample 8

Image 177: Portrait & Lifestyle sample 8

Image 178: Posters & Character Worlds sample 8

Image 179: Portrait & Lifestyle sample 9

Image 180: Concept & Imagination sample 9

Image 181: Concept & Imagination sample 10

Image 182: Portrait & Lifestyle sample 11

Image 183: Concept & Imagination sample 11

Image 184: Portrait & Lifestyle sample 12

Image 185: Posters & Character Worlds sample 12

Image 186: Information Design sample 14

Image 187: Posters & Character Worlds sample 14

Image 188: Portrait & Lifestyle sample 15

Image 189: Information Design sample 16

Image 190: Posters & Character Worlds sample 16

Image 191: Posters & Character Worlds sample 17

Image 192: Information Design sample 18

Image 193: Portrait & Lifestyle sample 20

Image 194: Portrait & Lifestyle sample 21

Image 195: Portrait & Lifestyle sample 22

Image 196: Information Design sample 24

Image 197: Information Design sample 26

Image 198: Portrait & Lifestyle sample 27

Image 199: Information Design sample 29

Image 200: Narrative Illustration sample 1

Image 201: Posters & Character Worlds sample 1

Image 202: Narrative Illustration sample 2

Image 203: Portrait & Lifestyle sample 3

Image 204: Narrative Illustration sample 3

Image 205: Concept & Imagination sample 3

Image 206: Portrait & Lifestyle sample 4

Image 207: Information Design sample 5

Image 208: Concept & Imagination sample 5

Image 209: Narrative Illustration sample 6

Image 210: Portrait & Lifestyle sample 7

Image 211: Concept & Imagination sample 7

Image 212: Narrative Illustration sample 8

Image 213: Information Design sample 9

Image 214: Narrative Illustration sample 9

Image 215: Information Design sample 10

Image 216: Narrative Illustration sample 10

Image 217: Posters & Character Worlds sample 10

Image 218: Posters & Character Worlds sample 11

Image 219: Narrative Illustration sample 12

Image 220: Portrait & Lifestyle sample 13

Image 221: Posters & Character Worlds sample 13

Image 222: Information Design sample 15

Image 223: Portrait & Lifestyle sample 16

Image 224: Information Design sample 17

Image 225: Portrait & Lifestyle sample 18

Image 226: Portrait & Lifestyle sample 19

Image 227: Information Design sample 20

Image 228: Information Design sample 22

Image 229: Portrait & Lifestyle sample 23

Image 230: Information Design sample 25

Image 231: Portrait & Lifestyle sample 26

Image 232: Information Design sample 28

Image 233: Information Design sample 30

Image 234: Narrative Illustration sample 1

Image 235: Posters & Character Worlds sample 1

Image 236: Narrative Illustration sample 2

Image 237: Portrait & Lifestyle sample 3

Image 238: Narrative Illustration sample 3

Image 239: Concept & Imagination sample 3

Image 240: Portrait & Lifestyle sample 4

Image 241: Information Design sample 5

Image 242: Concept & Imagination sample 5

Image 243: Narrative Illustration sample 6

Image 244: Portrait & Lifestyle sample 7

Image 245: Concept & Imagination sample 7

Image 246: Narrative Illustration sample 8

Image 247: Information Design sample 9

Image 248: Narrative Illustration sample 9

Image 249: Information Design sample 10

Image 250: Narrative Illustration sample 10

Image 251: Posters & Character Worlds sample 10

Image 252: Posters & Character Worlds sample 11

Image 253: Narrative Illustration sample 12

Image 254: Portrait & Lifestyle sample 13

Image 255: Posters & Character Worlds sample 13

Image 256: Information Design sample 15

Image 257: Portrait & Lifestyle sample 16

Image 258: Information Design sample 17

Image 259: Portrait & Lifestyle sample 18

Image 260: Portrait & Lifestyle sample 19

Image 261: Information Design sample 20

Image 262: Information Design sample 22

Image 263: Portrait & Lifestyle sample 23

Image 264: Information Design sample 25

Image 265: Portrait & Lifestyle sample 26

Image 266: Information Design sample 28

Image 267: Information Design sample 30

Image 268: Narrative Illustration sample 1

Image 269: Posters & Character Worlds sample 1

Image 270: Narrative Illustration sample 2

Image 271: Portrait & Lifestyle sample 3

Image 272: Narrative Illustration sample 3

Image 273: Concept & Imagination sample 3

Image 274: Portrait & Lifestyle sample 4

Image 275: Information Design sample 5

Image 276: Concept & Imagination sample 5

Image 277: Narrative Illustration sample 6

Image 278: Portrait & Lifestyle sample 7

Image 279: Concept & Imagination sample 7

Image 280: Narrative Illustration sample 8

Image 281: Information Design sample 9

Image 282: Narrative Illustration sample 9

Image 283: Information Design sample 10

Image 284: Narrative Illustration sample 10

Image 285: Posters & Character Worlds sample 10

Image 286: Posters & Character Worlds sample 11

Image 287: Narrative Illustration sample 12

Image 288: Portrait & Lifestyle sample 13

Image 289: Posters & Character Worlds sample 13

Image 290: Information Design sample 15

Image 291: Portrait & Lifestyle sample 16

Image 292: Information Design sample 17

Image 293: Portrait & Lifestyle sample 18

Image 294: Portrait & Lifestyle sample 19

Image 295: Information Design sample 20

Image 296: Information Design sample 22

Image 297: Portrait & Lifestyle sample 23

Image 298: Information Design sample 25

Image 299: Portrait & Lifestyle sample 26

Image 300: Information Design sample 28

Image 301: Information Design sample 30

The images above are shown solely to demonstrate model capabilities and do not imply any commercial use or endorsement. The model is intended for non-commercial personal derivative creation.

Posters and Design Layouts

ERNIE-Image excels at posters and other layout-critical images, combining strong visual focal points with reliable title placement, supporting text, and overall composition. Where many open weights models can produce a "poster-like" image but fall apart on typography or layout control, ERNIE-Image gets much closer to a usable design draft.

Anime and Storyboards

ERNIE-Image handles anime-style and storyboard generation well — the challenge here goes beyond drawing appealing characters to maintaining expressions, action clarity, scene transitions, and panel-level organization. Its outputs read more like actual manga pages than disconnected illustrations, making it a good fit for narrative visual content.

Multi-panel Visual Expression

Beyond single images, ERNIE-Image can produce grouped and sequential visuals — multi-panel compositions, meme-style image sets, and emotionally coherent grouped outputs. This cross-frame consistency is valuable when a task calls for pacing, progression, or repeated visual motifs rather than one isolated image.

Multilingual Text Rendering

ERNIE-Image is especially good at rendering readable, faithful multilingual text in visually complex images — both Chinese and English, covering dense paragraphs, layout-sensitive titles, callout annotations, anime dialogue, and poster copy. In practice, text quality often decides whether an image is merely plausible or genuinely usable, which is what makes this one of the model's most consequential capabilities.

Stylized and Film-like Aesthetics

Beyond clean graphic outputs, ERNIE-Image handles distinctive stylized aesthetics well — softer cinematic tones, film-like grain, and mood-driven visual expression. Many strong image models converge toward a similar high-saturation "AI look"; ERNIE-Image covers a broader aesthetic range with more recognizable stylistic character.

3. Model Results

We evaluate ERNIE-Image on four widely-used text-to-image generation benchmarks: GenEval (compositional generation), OneIG-EN / OneIG-ZH (open-domain image generation in English and Chinese), and LongTextBench (long-text rendering fidelity).

OneIG-EN

model Open Weights Alignment Text Reasoning Style Diversity Overall ↑
Nano Banana 2.0🥇 – 0.8880 0.9440 0.3340 0.4810 0.2450 0.5780
Seedream 4.5🥈 – 0.8910 0.9980 0.3500 0.4340 0.2070 0.5760
Image 302: ERNIEERNIE-Image (w/ PE)🥉 ✓ 0.8678 0.9788 0.3566 0.4309 0.2411 0.5750
Seedream 4.0 – 0.8920 0.9830 0.3470 0.4530 0.1910 0.5730
Image 303: ERNIEERNIE-Image-Turbo (w/ PE) ✓ 0.8676 0.9666 0.3537 0.4191 0.2212 0.5656
Image 304: ERNIEERNIE-Image (w/o PE) ✓ 0.8909 0.9668 0.2950 0.4471 0.1687 0.5537
Z-Image ✓ 0.8810 0.9870 0.2800 0.3870 0.1940 0.5460
Qwen-Image ✓ 0.8820 0.8910 0.3060 0.4180 0.1970 0.5390
Image 305: ERNIEERNIE-Image-Turbo (w/o PE) ✓ 0.8795 0.9488 0.2913 0.4277 0.1232 0.5341
GPT Image 1 [High] – 0.8510 0.8570 0.3450 0.4620 0.1510 0.5330
FLUX.2-klein-9B ✓ 0.8871 0.8657 0.3117 0.4417 0.1560 0.5324
Seedream 3.0 – 0.8180 0.8650 0.2750 0.4130 0.2770 0.5300
Qwen-Image-2512 ✓ 0.8760 0.9900 0.2920 0.3380 0.1510 0.5300
GLM-Image ✓ 0.8050 0.9690 0.2980 0.3530 0.2130 0.5280
Z-Image-Turbo ✓ 0.8400 0.9940 0.2980 0.3680 0.1390 0.5280
HiDream-I1-Full ✓ 0.8290 0.7070 0.3170 0.3470 0.1860 0.4770
Kolors 2.0 ✓ 0.8200 0.4270 0.2620 0.3600 0.3000 0.4340
BAGEL ✓ 0.7690 0.2440 0.1730 0.3670 0.2510 0.3610
BLIP3-o ✓ 0.7110 0.1330 0.2230 0.3610 0.2290 0.3070
Janus-Pro ✓ 0.5530 0.0010 0.1390 0.2760 0.3650 0.2670

OneIG-ZH

model Open Weights Alignment Text Reasoning Style Diversity Overall ↑
Nano Banana 2.0🥇 – 0.8430 0.9830 0.3110 0.4610 0.2360 0.5670
Image 306: ERNIEERNIE-Image (w/ PE)🥈 ✓ 0.8299 0.9539 0.3056 0.4342 0.2478 0.5543
Seedream 4.0🥉 – 0.8360 0.9860 0.3040 0.4430 0.2000 0.5540
Seedream 4.5 – 0.8320 0.9860 0.3000 0.4260 0.2130 0.5510
Qwen-Image ✓ 0.8250 0.9630 0.2670 0.4050 0.2790 0.5480
Image 307: ERNIEERNIE-Image-Turbo (w/ PE) ✓ 0.8258 0.9386 0.3043 0.4208 0.2281 0.5435
Z-Image ✓ 0.7930 0.9880 0.2660 0.3860 0.2430 0.5350
Seedream 3.0 – 0.7930 0.9280 0.2810 0.3970 0.2430 0.5280
Image 308: ERNIEERNIE-Image (w/o PE) ✓ 0.8421 0.8979 0.2656 0.4212 0.1772 0.5208
Qwen-Image-2512 ✓ 0.8230 0.9830 0.2720 0.3420 0.1570 0.5150
GLM-Image ✓ 0.7380 0.9760 0.2840 0.3350 0.2210 0.5110
Z-Image-Turbo ✓ 0.7820 0.9820 0.2760 0.3610 0.1340 0.5070
Image 309: ERNIEERNIE-Image-Turbo (w/o PE) ✓ 0.8326 0.9086 0.2580 0.4002 0.1316 0.5062
GPT Image 1 [High] – 0.8120 0.6500 0.3000 0.4490 0.1590 0.4740
FLUX.2-klein-9B ✓ 0.8201 0.4920 0.2599 0.4166 0.1625 0.4302
Kolors 2.0 ✓ 0.7380 0.5020 0.2260 0.3310 0.3330 0.4260
BAGEL ✓ 0.6720 0.3650 0.1860 0.3570 0.2680 0.3700
HiDream-I1-Full ✓ 0.6200 0.2050 0.2560 0.3040 0.3000 0.3370
BLIP3-o ✓ 0.6080 0.0920 0.2130 0.3690 0.2330 0.3030
Janus-Pro ✓ 0.3240 0.1480 0.1040 0.2640 0.3580 0.2400

LongTextBench

model Open Weights LongText-Bench-EN LongText-Bench-ZH Overall ↑
Seedream 4.5🥇 – 0.9890 0.9873 0.9882
Image 310: ERNIEERNIE-Image (w/ PE)🥈 ✓ 0.9804 0.9661 0.9733
GLM-Image🥉 ✓ 0.9524 0.9788 0.9656
Image 311: ERNIEERNIE-Image-Turbo (w/ PE) ✓ 0.9675 0.9636 0.9655
Nano Banana 2.0 – 0.9808 0.9491 0.9650
Image 312: ERNIEERNIE-Image-Turbo (w/o PE) ✓ 0.9602 0.9675 0.9639
Image 313: ERNIEERNIE-Image (w/o PE) ✓ 0.9679 0.9594 0.9636
Qwen-Image-2512 ✓ 0.9561 0.9647 0.9604
Qwen-Image ✓ 0.9430 0.9460 0.9445
Z-Image ✓ 0.9350 0.9360 0.9355
Seedream 4.0 – 0.9214 0.9261 0.9238
Z-Image-Turbo ✓ 0.9170 0.9260 0.9215
Seedream 3.0 – 0.8960 0.8780 0.8870
GPT Image 1 [High] – 0.9560 0.6190 0.7875
FLUX.2-klein-9B ✓ 0.8642 0.2183 0.5413
BAGEL ✓ 0.3730 0.3100 0.3415
Kolors 2.0 ✓ 0.2580 0.3290 0.2935
HiDream-I1-Full ✓ 0.5430 0.0240 0.2835
BLIP3-o ✓ 0.0210 0.0180 0.0195
Janus-Pro ✓ 0.0190 0.0060 0.0125

GenEval

model Open Weights Single Object Two Object Counting Colors Position Attribute Binding Overall ↑
Image 314: ERNIEERNIE-Image (w/o PE)🥇 ✓ 1.0000 0.9596 0.7781 0.9282 0.8550 0.7925 0.8856
Image 315: ERNIEERNIE-Image (w/ PE)🥈 ✓ 0.9906 0.9596 0.8187 0.8830 0.8625 0.7225 0.8728
Qwen-Image🥉 ✓ 0.9900 0.9200 0.8900 0.8800 0.7600 0.7700 0.8683
Image 316: ERNIEERNIE-Image-Turbo (w/o PE) ✓ 1.0000 0.9621 0.7906 0.9202 0.7975 0.7300 0.8667
Image 317: ERNIEERNIE-Image-Turbo (w/ PE) ✓ 0.9938 0.9419 0.8375 0.8351 0.7950 0.7025 0.8510
FLUX.2-klein-9B ✓ 0.9313 0.9571 0.8281 0.9149 0.7175 0.7400 0.8481
Seedream 3.0 – 0.9900 0.9600 0.9100 0.9300 0.4700 0.8000 0.8433
Z-Image ✓ 1.0000 0.9400 0.7800 0.9300 0.6200 0.7700 0.8400
GPT Image 1 – 0.9900 0.9200 0.8500 0.9200 0.7500 0.6100 0.8400
HiDream-I1-Full ✓ 1.0000 0.9800 0.7900 0.9100 0.6000 0.7200 0.8333
Z-Image-Turbo ✓ 1.0000 0.9500 0.7700 0.8900 0.6500 0.6800 0.8233
Janus-Pro-7B ✓ 0.9900 0.8900 0.5900 0.9000 0.7900 0.6600 0.8000
SD3.5-Large ✓ 0.9800 0.8900 0.7300 0.8300 0.3400 0.4700 0.7100
FLUX.1 [Dev] ✓ 0.9800 0.8100 0.7400 0.7900 0.2200 0.4500 0.6600
JanusFlow ✓ 0.9700 0.5900 0.4500 0.8300 0.5300 0.4200 0.6300
SD3 Medium ✓ 0.9800 0.7400 0.6300 0.6700 0.3400 0.3600 0.6200
Emu3-Gen ✓ 0.9800 0.7100 0.3400 0.8100 0.1700 0.2100 0.5400
Show-o ✓ 0.9500 0.5200 0.4900 0.8200 0.1100 0.2800 0.5300
PixArt-α ✓ 0.9800 0.5000 0.4400 0.8000 0.0800 0.0700 0.4800

Key Findings

  • •Competitive with the best. ERNIE-Image consistently places in the top 3 across all four benchmarks: #1 on GenEval (0.8856), #2 on OneIG-ZH (0.5543) and LongTextBench (0.9733), and #3 on OneIG-EN (0.5750), going head-to-head with leading closed-source models such as Nano Banana 2.0 and Seedream 4.5.
  • •Strongest open weights model. ERNIE-Image ranks #1 among all open weights models on every benchmark by a clear margin. These results come from just an 8B-parameter DiT backbone — one of the most parameter-efficient architectures at this performance level.
  • •Excellent text rendering. ERNIE-Image ranks #2 on LongTextBench with strong performance in both English and Chinese long-text rendering, and posts highly competitive Text scores on both OneIG-EN and OneIG-ZH — underscoring its edge in multilingual text generation.

4. Prompt Enhancer (PE)

ERNIE-Image works best with long, detailed, well-structured prompts — richer descriptions tend to produce better generation quality, tighter instruction fidelity, and more faithful rendering of complex layouts or narrative content. In practice, though, users typically type a short sentence rather than the kind of detailed prompt that plays to the model's strengths.

To bridge this gap, we release a built-in 3B Prompt Enhancer that expands short user inputs into more detailed, structured prompts better suited to ERNIE-Image. The goal is not to change the user's intent, but to translate concise requests into a form that gets more out of the model — especially for posters, anime, webpage layouts, game screenshots, and other structured visual tasks.

The examples below illustrate this effect. Without prompt enhancement, the model tends to interpret short prompts literally and incompletely. With our 3B Prompt Enhancer, prompts become more descriptive and structured, noticeably improving results in many scenarios. We also find that stronger LLMs can push this even further — a sign that prompt enhancement is a practical lever for getting more out of ERNIE-Image's long-prompt generation ability.

Solve AIME Problem Comic Strip Storytelling HD-2D RPG Screenshot LLM Webpage

Source Prompt

Solve this problem on the whiteboard: Find the number of positive integer palindromes written in base 10 with no zero digits, whose digits sum to 13. For example, 42124 has these properties. Recall that a palindrome is a number whose representation reads the same from left to right as from right to left.

w/o pe

Image 318: w/o pe

w/ pe Our 3B Model

Image 319: w/ pe

w/ pe (LLMs)gemini-3.1-pro-preview

Image 320: w/ pe (LLMs)

Analysis

The source prompt asks the model to solve an AIME problem on a whiteboard. Without PE, the model simply copies the problem statement verbatim — with multiple spelling errors (e.g. "interger", "palindrimes") — and makes no attempt at a solution. With the 3B PE, a structured solution layout emerges — step decomposition, palindrome representation, combination enumeration — but the final answer is wrong (24; correct: 62), limited by the small model's reasoning capacity. With LLM enhancement, the whiteboard shows a complete, rigorous derivation: parity analysis, central-digit enumeration, the composition theorem, and the correct answer of 62, rendered in multi-color with professional formatting.

w/o pe

Without Prompt Enhancement

Given a brief prompt, ERNIE-Image tends toward a literal, surface-level interpretation — it faithfully renders the elements mentioned but won't inject domain knowledge, refine the layout, or fill in missing detail on its own. For tasks that need knowledge or complex structural planning, this literal reading falls short.

w/ pe

With Our 3B Prompt Enhancer

We release a lightweight prompt enhancement model based on a fine-tuned Ministral 3B. It expands brief user prompts into detailed, structured descriptions, noticeably lifting generation quality in many scenarios. At 3B parameters, though, it can fall short on tasks that demand deep domain knowledge (e.g., mathematical reasoning) or exceptionally rich world knowledge.

w/ pe (LLMs)

With Large LLM Enhancement

We find that using an LLM such as gemini-3.1-pro-preview for zero-shot prompt enhancement yields markedly better results across the board. With deeper world knowledge and stronger reasoning, the LLM produces information-dense, structurally precise prompts that bring out ERNIE-Image's full generation potential.

5. Conclusion

ERNIE-Image shows that a compact 8B model can rival much larger open weights text-to-image systems while staying practical for real-world use. Across the examples and benchmarks in this blog, the model demonstrates particular strengths in text rendering, complex instruction following, structured visual integrity, and stylistically diverse image generation. These capabilities matter most for high-value tasks like posters, anime, multi-panel layouts, and other scenarios where both visual quality and controllability count. We hope the released ERNIE-Image and ERNIE-Image-Turbo checkpoints serve as a useful open foundation for research, development, and creative applications.

Get Started with ERNIE-Image

ERNIE-Image Main model on HuggingFace Learn More#### ERNIE-Image-Turbo Fast 8-step distilled version Learn More#### Prompt Gallery Browse curated generation examples Learn More

Copyright © 2025 Baidu, Inc. All Rights Reserved.

來源:百度文心 ERNIE 官方博客 · ernie.baidu.com