Text-to-Image Retrieval on FlowVQA hard negatives (test)
89.973R@1GRAN - SaCLIP
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GRAN - SaCLIPData preparation strategy=GRAN, Finetuning strategy=SaCLIP2026.02 | 89.973 | 98.602 | 94.133 | |
| PSEUDO - SaCLIPData preparation strategy=PSEUDO, Finetuning strategy=SaCLIP2026.02 | 70.484 | 87.204 | 79.069 | |
| GRAN - CLIPData preparation strategy=GRAN, Finetuning strategy=CLIP2026.02 | 61.64 | 82.634 | 72.529 | |
| w/o FTData preparation strategy=w/o FT2026.02 | 60.538 | 82.419 | 71.632 | |
| PSEUDO - CLIPData preparation strategy=PSEUDO, Finetuning strategy=CLIP2026.02 | 50.511 | 76.022 | 63.644 | |
| CROP - CLIPData preparation strategy=CROP, Finetuning strategy=CLIP2026.02 | 49.14 | 75.403 | 62.629 |