Image generation
T2I-CoReBench
Text-to-image composition and visual reasoning, evaluated by T2I-CoReBench.
Published by T2I-CoReBench under CC BY-SA 4.0. Compare it with other tests
Top 15
Each model at its best configuration, coloured by maker.
The best model from each maker
Each maker’s highest-scoring model on this test, and its rank.
| Maker | Best model | Score | Rank |
|---|---|---|---|
 Google | Nano Banana 2 | 85.3% | #1 of 40 |
 OpenAI | gpt-image-1.5 | 78.2% | #3 of 40 |
 ByteDance | Seedream 4.5 | 77.0% | #4 of 40 |
 Black Forest Labs | FLUX.2 Dev | 64.4% | #10 of 40 |
 Alibaba | Qwen Image 2512 | 62.4% | #11 of 40 |
 Tencent | HunyuanImage 3.0 | 58.7% | #16 of 40 |
 Stability AI | SD 3.5 Large | 37.8% | #29 of 40 |
Every result
40 models, each once at its best configuration, highest score first. Models with the same score share a rank.
| Rank | Model | Maker | Score |
|---|---|---|---|
| #1 | Nano Banana 2 | Google | 85.3% |
| #2 | Nano-Banana-Pro | Google | 85.0% |
| #3 | gpt-image-1.5 | OpenAI | 78.2% |
| #4 | Seedream 4.5 | ByteDance | 77.0% |
| #5 | Imagen-4-Ultra | Google | 76.1% |
| #6 | Nano-Banana | Google | 75.9% |
| #7 | Seedream 4.0 | ByteDance | 75.3% |
| #8 | GPT Image | OpenAI | 72.6% |
| #9 | Imagen-4 | Google | 70.0% |
| #10 | FLUX.2 Dev | Black Forest Labs | 64.4% |
| #11 | Qwen Image 2512 | Alibaba | 62.4% |
| #12 | Z Image |  Not reported | 61.1% |
| #13 | FLUX.2 [klein] 9B | Black Forest Labs | 60.6% |
| #14 | LongCat Image | Not reported | 59.6% |
| #15 | Qwen Image | Alibaba | 58.9% |
| #16 | HunyuanImage 3.0 | Tencent | 58.7% |
| #17 | Z Image Turbo | Not reported | 56.4% |
| #18 | Seedream 3.0 | ByteDance | 56.0% |
| #19 | Gemini-2.0-Flash | Google | 54.8% |
| #20 | FLUX.2 Klein 4B | Black Forest Labs | 54.6% |
| #21 | FLUX.1 Krea Dev | Black Forest Labs | 48.2% |
| #22 | Infinity 8B | Not reported | 43.6% |
| #23 | FLUX.1-schnell | Black Forest Labs | 43.2% |
| #24 | HiDream I1 | Not reported | 43.0% |
| #25 | OmniGen2 7B | Not reported | 42.9% |
How it is scored
Published overall scores with the Gemini 2.5 Flash evaluator. These measure composition and reasoning across 12 dimensions, not human aesthetic preference. Prompt-rewriting variants and other evaluators are excluded. Coverage is limited to models evaluated by the source.
Results from T2I-CoReBench (CC BY-SA 4.0), as published. Model names link to their pages where the site can match them; makers come from the model’s page, or else from the publisher.