Param-Retouch on Aether-Bench Param 1M+ (test)
0.023L1 LossOurs-SFT
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Ours-SFTReasoning capability=true2026.04 | 0.023 | 30.39 | 94.6 | 0.039 | 0.04 | 0.071 | 0.664 | |
| Ours-DAPO-AEReasoning capability=true2026.04 | 0.024 | 30.18 | 94.7 | 0.041 | 0.042 | 0.067 | 0.644 | |
| Flux.1 KontextReasoning capability=false2026.04 | 0.14 | 18.51 | 78.3 | 0.204 | 0.136 | 0.008 | 0.468 | |
| Qwen-Image-2509Reasoning capability=false2026.04 | 0.283 | 13.55 | 38 | 0.484 | 0.257 | 0.198 | 1.843 |