UI-to-Code Generation on Design2Code 50-page (val)
0.8127CLIP SimilarityUIPress-256 (E17)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| UIPress-256 (E17)Base model=Qwen3-VL-8B-Instruct, Type=optical, Tokens=256, Compression Ratio=25.5×, Evaluation Hardware=single A40 GPU2026.04 | 0.8127 | 0.786 | 42 | 90.4 | |
| Resolution scaling (480px)Base model=Qwen3-VL-8B-Instruct, Type=resolution, Tokens=845, Compression Ratio=7.7×, Evaluation Hardware=single A40 GPU2026.04 | 0.7768 | 0.749 | 69 | 94 | |
| Qwen3-VL (native)Base model=Qwen3-VL-8B-Instruct, Type=—, Tokens=6,517, Compression Ratio=1.0×, Evaluation Hardware=single A40 GPU2026.04 | 0.7563 | 0.71 | 384 | 124.1 | |
| FastV (75% zero)Base model=Qwen3-VL-8B-Instruct, Type=zeroing, Tokens=6,517†, Compression Ratio=1.0×, Evaluation Hardware=single A40 GPU2026.04 | 0.754 | 0.727 | 385 | 118.2 | |
| EfficientUI (60% prune)Base model=Qwen3-VL-8B-Instruct, Type=element, Tokens=730, Compression Ratio=8.9×, Evaluation Hardware=single A40 GPU2026.04 | 0.7523 | 0.724 | 99 | 99.5 | |
| VisionZip-256Base model=Qwen3-VL-8B-Instruct, Type=selection, Tokens=256, Compression Ratio=25.5×, Evaluation Hardware=single A40 GPU2026.04 | 0.7333 | 0.705 | 45 | 110.4 |