Screenshot-to-code generation on Design2Code HARD
80.8Block MatchVisRefiner-7B
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| VisRefiner-7BInference Mode=Self-refinement2026.02 | 80.8 | 96.2 | 82.2 | 84.8 | 87.7 | 86.3 | |
| VisRefiner-7BInference Mode=Single-step2026.02 | 78.1 | 96.2 | 82.3 | 83.7 | 87.9 | 85.6 | |
| Claude 3 OpusInference Mode=Self-refinement2026.02 | 73.3 | 95.9 | 76.6 | 70 | 85.6 | 80.3 | |
| Qwen2.5-VL-72BInference Mode=Self-refinement2026.02 | 72.8 | 94.9 | 79.8 | 80.9 | 84.9 | 82.7 | |
| GPT-4oInference Mode=Self-refinement2026.02 | 72.1 | 96.4 | 81.1 | 82.4 | 88.2 | 84 | |
| Claude 3.5 SonnetInference Mode=Self-refinement2026.02 | 71.9 | 96.5 | 82.6 | 83 | 88.8 | 84.6 | |
| Qwen2.5-VL-72BInference Mode=Single-step2026.02 | 64.9 | 94.8 | 79.2 | 80.4 | 84.8 | 80.8 | |
| Qwen2.5-VL-7BInference Mode=Self-refinement2026.02 | 63.2 | 92.6 | 74 | 76.1 | 83.2 | 77.8 | |
| MiMo-VL-7B-RLInference Mode=Self-refinement2026.02 | 62.3 | 91.9 | 77.1 | 77.9 | 84.4 | 78.7 | |
| Qwen2.5-VL-7BInference Mode=Single-step2026.02 | 62.2 | 91.4 | 73 | 74.3 | 82.8 | 76.7 | |
| Claude 3.5 SonnetInference Mode=Single-step2026.02 | 61.7 | 91.1 | 83 | 84.4 | 89.5 | 81.9 | |
| MiMo-VL-7B-RLInference Mode=Single-step2026.02 | 61.1 | 91 | 77.8 | 77.5 | 84.5 | 78.4 | |
| Claude 3 OpusInference Mode=Single-step2026.02 | 57.1 | 88.7 | 74.2 | 72.4 | 85.8 | 75.6 | |
| GPT-4oInference Mode=Single-step2026.02 | 56.6 | 89.8 | 78.6 | 81.9 | 87.1 | 78.8 |