Reasoning-based Image Editing on UniREditBench 44 (test)
81Real World ScoreGPT-Image-1
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GPT-Image-1Params=Not specified2026.02 | 81 | 62.1 | 73.4 | |
| Nano BananaParams=Not specified2026.02 | 75.2 | 60.4 | 68.3 | |
| DeepGen 1.0 (SFT)Params=3B + 2B, training_mode=SFT2026.02 | 74.3 | 80.7 | 77.5 | |
| DeepGen 1.0 (RL)Params=3B + 2B, training_mode=RL2026.02 | 73.2 | 78.2 | 75.7 | |
| Qwen-Image edit [2509]Params=7B + 20B2026.02 | 71 | 41.9 | 56.5 | |
| Seedream 4.0Params=Not specified2026.02 | 66.2 | 45.4 | 55.8 | |
| BAGELParams=14B, CoT reasoning=true2026.02 | 56.8 | 45.1 | 51 | |
| OmniGen2Params=3B + 4B2026.02 | 53.7 | 33.1 | 43.4 | |
| Lumina-DiMOOParams=8B2026.02 | 51.4 | 45.6 | 48.5 | |
| FLUX-Kontext-ProParams=Not specified2026.02 | 45 | 46.5 | 45.8 |