Dense GUI Parsing on GroundCUA full benchmark
39.8Page IoUOmniParser V2
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| OmniParser V2Size=20M, Training=Fine-tuned on ScreenParse2026.02 | 39.8 | 6.1 | |
| RT-DETRv2Size=43M, Training=Trained on ScreenParse2026.02 | 38.8 | 5.9 | |
| YOLO (YOLOv11-large)Size=25M, Training=Trained on ScreenParse2026.02 | 37.9 | 5.7 | |
| OmniParser V2Size=20M, Status=Off-the-shelf2026.02 | 36.1 | 4.9 | |
| ScreenVLMSize=316M2026.02 | 25.1 | 4.3 | |
| InternVL3-2BSize=2B, Training=Fine-tuned on ScreenParse2026.02 | 20.3 | 3.6 | |
| Qwen3-VL-2B-InstructSize=2B, Training=Fine-tuned on ScreenParse2026.02 | 9 | 1.9 | |
| Qwen3-VL-8B-InstructSize=8B, Protocol=Prompted2026.02 | 6 | 1 | |
| Qwen3-VL-2B-InstructSize=2B, Protocol=Prompted2026.02 | 3 | 0.5 | |
| InternVL3-2BSize=2B, Protocol=Prompted2026.02 | 2.5 | 0.6 |