Object Replacement on COCO subset (test)
59S2 ScoreVISTA
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| VISTAVisual Encoder=DINOv2, LLM Backbone=Gemma-2-2B-it2026.05 | 59 | 28 | 13 | 1.46 | |
| PatchSAEVisual Encoder=CLIP Local2026.05 | 34 | 31 | 35 | 0.99 | |
| VisionSAEVisual Encoder=CLIP Local2026.05 | 23 | 14 | 63 | 0.6 | |
| VisionSAEVisual Encoder=CLIP Local (Rem+Rep)2026.05 | 22 | 2 | 76 | 0.46 | |
| Ablation (No SAE)Visual Encoder=I-JEPA2026.05 | 20 | 9 | 71 | 0.49 | |
| VISTAVisual Encoder=I-JEPA - Gemma 2B, LLM Backbone=Gemma-2-2B-it2026.05 | 17 | 47 | 36 | 0.81 | |
| Ablation (No SAE)Visual Encoder=CLIP2026.05 | 16 | 6 | 78 | 0.38 | |
| VisionSAEVisual Encoder=CLIP Global2026.05 | 8 | 49 | 43 | 0.65 | |
| VISTAVisual Encoder=CLIP - Gemma 2B, LLM Backbone=Gemma-2-2B-it2026.05 | 7 | 4 | 89 | 0.18 | |
| Ablation (No SAE)Visual Encoder=DINOv22026.05 | 3 | 0 | 97 | 0.06 | |
| VisionSAEVisual Encoder=CLIP Global (Rem+Rep)2026.05 | 0 | 1 | 99 | 0.01 |