Style-Retouch on Aether-Bench Style 1M+ (test)
0.092L1 LossOurs-DAPO-AE
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Ours-DAPO-AEReasoning capability=true2026.04 | 0.092 | 20.12 | 0.847 | 0.145 | 0.099 | 0.036 | 0.526 | |
| Flux.1 KontextReasoning capability=false2026.04 | 0.094 | 19.48 | 0.831 | 0.162 | 0.106 | 0.048 | 0.741 | |
| Ours-SFTReasoning capability=true2026.04 | 0.097 | 19.73 | 0.839 | 0.149 | 0.1 | 0.039 | 0.592 | |
| Nano BananaReasoning capability=true2026.04 | 0.125 | 16.66 | 0.596 | 0.242 | 0.138 | 0.151 | 1.579 | |
| JarvisArtReasoning capability=true2026.04 | 0.147 | 15.72 | 0.677 | 0.288 | 0.17 | 0.1 | 1.235 | |
| Qwen-Image-2509Reasoning capability=false2026.04 | 0.158 | 14.34 | 0.494 | 0.289 | 0.196 | 0.174 | 1.725 |