Screenshot-to-code generation on VisDiffUI (test)
96.6Block MatchVisRefiner-7B
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| VisRefiner-7BInference Mode=Self-refinement2026.02 | 96.6 | 99.1 | 87.3 | 89.7 | 90 | 92.5 | |
| Qwen2.5-VL-72BInference Mode=Self-refinement2026.02 | 95.4 | 99 | 87.3 | 86.1 | 88.6 | 91.3 | |
| Qwen2.5-VL-72BInference Mode=Single-step2026.02 | 95.2 | 99 | 87.4 | 85.4 | 88.1 | 91 | |
| VisRefiner-7BInference Mode=Single-step2026.02 | 93.9 | 99.3 | 86.9 | 89.3 | 90.5 | 92 | |
| Qwen2.5-VL-7BInference Mode=Single-step2026.02 | 91.9 | 98.8 | 80.3 | 79.2 | 87.2 | 87.5 | |
| Qwen2.5-VL-7BInference Mode=Self-refinement2026.02 | 91.5 | 97.9 | 79.6 | 78.1 | 87 | 86.8 | |
| MiMo-VL-7B-RLInference Mode=Self-refinement2026.02 | 87.4 | 94.5 | 79.7 | 76.3 | 85.9 | 84.8 | |
| MiMo-VL-7B-RLInference Mode=Single-step2026.02 | 85.6 | 94.9 | 81.6 | 77.7 | 86.3 | 85.2 |