Action-conditioned image generation evaluation on Do-Undo zero-shot 1.0
8.44Instruction Adherence (IF)Qwen Image
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen Imagejudge=Gemini 22025.12 | 8.44 | 9.35 | 8.15 | 9.18 | |
| Do-Undojudge=Gemini 22025.12 | 8.05 | 9.12 | 8.03 | 9.06 | |
| Do-Undo(c)judge=Gemini 2, configuration=canonical2025.12 | 8.02 | 9.11 | 7.94 | 9.09 | |
| Geminitype=Und&Gen2025.12 | 8.01 | 8.3 | — | 8.39 | |
| Qwen (20B)type=Und&Gen2025.12 | 8 | 7.64 | — | 7.82 | |
| Bageljudge=Gemini 22025.12 | 7.97 | 9.3 | 8.17 | 9.26 | |
| Bagel-think (7B)type=Und&Gen2025.12 | 7.62 | 7.57 | — | 7.79 | |
| Do-Undotype=Und&Gen2025.12 | 7.52 | 7.5 | — | 7.61 | |
| FluxKontext (7B)type=Gen2025.12 | 7.04 | 6.61 | — | 6.74 | |
| FluxKontextjudge=Gemini 22025.12 | 6.69 | 8.82 | 6.91 | 8.81 |