Interactive Image Editing on CanvasCraft-RL (evaluation)
0.869Alignment ScoreCanvasAgent (SFT+RL)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| CanvasAgent (SFT+RL)Training Strategy=SFT+RL2026.07 | 0.869 | 0.821 | 0.762 | 0.849 | 0.785 | 5.436 | |
| GPT-Image-2Model Type=Image-only reference model2026.07 | 0.799 | — | 0.895 | — | — | — | |
| CanvasAgent (SFT)Training Strategy=SFT2026.07 | 0.613 | 0.557 | 0.711 | 0.576 | 0.467 | 1.32 | |
| Wan2.7-ImageModel Type=Image-only reference model2026.07 | 0.605 | — | 0.843 | — | — | — | |
| Qwen-Image-2.0Model Type=Image-only reference model2026.07 | 0.543 | — | 0.825 | — | — | — | |
| Qwen3-VL-8B-InstructModel Scale=8B2026.07 | 0.493 | 0.426 | 0.667 | 0.092 | 0.483 | 1.488 | |
| LLaVA-OneVision-7BModel Scale=7B2026.07 | 0.484 | 0.402 | 0.598 | 0.132 | 0.427 | 1.354 | |
| Qwen3-VL-32B-InstructModel Scale=32B2026.07 | 0.428 | 0.474 | 0.588 | 0.512 | 0.461 | 7.668 |