Industrial Anomaly Detection on MMVIAD Unseen
88.5Defect Detection AccuracyHuman (expert)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Human (expert)Type=-, Venue=-, Param.=-2026.05 | 88.5 | 78.9 | 80.1 | 96.4 | 86 | |
| Human (ordinary)Type=-, Venue=-, Param.=-2026.05 | 83.2 | 73.3 | 77.8 | 93.6 | 82 | |
| Gemini 3.1 ProType=Commercial, Venue=-, Param.=-2026.05 | 69.1 | 46.9 | 59.3 | 80.3 | 63.9 | |
| GPT-5.4-miniType=Commercial, Venue=-, Param.=-2026.05 | 62 | 29.8 | 60 | 62.2 | 53.5 | |
| VISTA (Qwen3-VL)Type=Ours, Venue=-, Param.=8B2026.05 | 60.7 | 37.6 | 49.6 | 81.9 | 57.5 | |
| GPT-5.4Type=Commercial, Venue=-, Param.=-2026.05 | 58.2 | 38 | 52 | 65.7 | 53.5 | |
| Video-R1Type=Open Source, Venue=NeurIPS’25, Param.=7B2026.05 | 49.5 | 31.1 | 48.4 | 57 | 46.5 | |
| Time-R1Type=Open Source, Venue=NeurIPS’25, Param.=7B2026.05 | 44.2 | 35.6 | 48.6 | 56.4 | 46.2 | |
| VideoChat-R1.5Type=Open Source, Venue=NeurIPS’25, Param.=7B2026.05 | 41.7 | 34.8 | 45.4 | 55.4 | 44.3 | |
| VideoChat-R1Type=Open Source, Venue=arXiv’25, Param.=7B2026.05 | 41.1 | 34.9 | 48.8 | 56.9 | 45.4 | |
| Qwen3-VLType=Open Source, Venue=arXiv’25, Param.=8B2026.05 | 39.5 | 36.2 | 37 | 67.4 | 45 | |
| Qwen3-VLType=Open Source, Venue=arXiv’25, Param.=4B2026.05 | 35.2 | 33.5 | 37.8 | 56.9 | 40.9 | |
| Qwen3-VLType=Open Source, Venue=arXiv’25, Param.=2B2026.05 | 32.3 | 26.2 | 27.4 | 54.8 | 35.2 |