Attention-only performance evaluation on CLIP
2.15SpeedupELSA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ELSAModel=ViT-L/14-336, Resolution=560, Precision=FP32, Scope=attention-only proxy2026.04 | 2.15 | 39.6 | |
| ELSAModel=ViT-L/14, Resolution=560, Precision=FP32, Scope=attention-only proxy2026.04 | 2.12 | 39.6 | |
| ELSAModel=ViT-B/16, Resolution=560, Precision=FP32, Scope=attention-only proxy2026.04 | 1.98 | 36.1 | |
| ELSAModel=ViT-L/14-336, Resolution=448, Precision=FP32, Scope=attention-only proxy2026.04 | 1.9 | 20.8 | |
| ELSAModel=ViT-L/14, Resolution=448, Precision=FP32, Scope=attention-only proxy2026.04 | 1.87 | 20.8 | |
| ELSAModel=ViT-B/16, Resolution=448, Precision=FP32, Scope=attention-only proxy2026.04 | 1.79 | 17.8 | |
| ELSAModel=ViT-L/14-336, Resolution=336, Precision=FP32, Scope=attention-only proxy2026.04 | 1.71 | 5 | |
| ELSAModel=ViT-L/14, Resolution=336, Precision=FP32, Scope=attention-only proxy2026.04 | 1.68 | 5 | |
| ELSAModel=ViT-B/16, Resolution=336, Precision=FP32, Scope=attention-only proxy2026.04 | 1.63 | 2.3 | |
| ELSAModel=ViT-L/14-336, Resolution=224, Precision=FP32, Scope=attention-only proxy2026.04 | 1.56 | 0 | |
| ELSAModel=ViT-L/14, Resolution=224, Precision=FP32, Scope=attention-only proxy2026.04 | 1.54 | 0 | |
| ELSAModel=ViT-B/16, Resolution=224, Precision=FP32, Scope=attention-only proxy2026.04 | 1.46 | 0 | |
| ELSAModel=ViT-B/32, Resolution=448, Precision=FP32, Scope=attention-only proxy2026.04 | 1.42 | 0 | |
| ELSAModel=ViT-B/32, Resolution=224, Precision=FP32, Scope=attention-only proxy2026.04 | 1.03 | 0 |