Compositional Evaluation on SugarCrepe swap att (test)
82.1AccuracyDeGLA
Evaluation Results
| Method | Links | |
|---|---|---|
| DeGLAYear=20252025.12 | 82.1 | |
| ABE-FG-CLIPBackbone=ViT-B/162025.12 | 81.98 | |
| TFLocalBackbone=PE, Training data=TROHN-Img2026.04 | 80.8 | |
| ABE-FG-CLIPBackbone=ViT-L/142025.12 | 80.78 | |
| TFLocalBackbone=PE, Training data=COCO2026.04 | 80.6 | |
| Structure-CLIPYear=20242025.12 | 80.5 | |
| FTBackbone=PE, Training data=TROHN-Img2026.04 | 80.3 | |
| TFLocalBackbone=CLIP, Training data=TROHN-Img2026.04 | 77.1 | |
| CE-CLIPYear=20242025.12 | 77 | |
| BaseBackbone=PE, Training data=Pretrained2026.04 | 76.8 | |
| TFLocalBackbone=CLIP, Training data=COCO2026.04 | 76.3 | |
| FTBackbone=PE, Training data=COCO2026.04 | 76.1 | |
| NegCLIPYear=20232025.12 | 75.23 | |
| TFGlobalBackbone=PE, Training data=TROHN-Img2026.04 | 74.9 | |
| LABCLIPYear=20252025.12 | 74.62 | |
| FTBackbone=CLIP, Training data=TROHN-Img2026.04 | 72 | |
| FG-CLIPBackbone=ViT-B/162025.12 | 71.02 | |
| FG-CLIPBackbone=ViT-L/142025.12 | 69.82 | |
| FTBackbone=CLIP, Training data=COCO2026.04 | 68.9 | |
| TFGlobalBackbone=CLIP, Training data=TROHN-Img2026.04 | 68.8 | |
| TFGlobalBackbone=PE, Training data=COCO2026.04 | 68 | |
| CLIPBackbone=ViT-B/162025.12 | 64.86 | |
| ABE-CLIPBackbone=ViT-B/162025.12 | 64.86 | |
| TFGlobalBackbone=CLIP, Training data=COCO2026.04 | 63.5 | |
| ABE-CLIPBackbone=ViT-L/142025.12 | 63.36 | |
| BaseBackbone=CLIP, Training data=Pretrained2026.04 | 63.3 | |
| CLIPBackbone=ViT-L/142025.12 | 62.46 |