Image-Language Understanding on COCO (Score, Retention Rate)
1ScoreVanilla
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VanillaTokens Retained=Original (2880), Compression Rate=0.0%2026.01 | 1 | 100 | |
| ViTCoPTokens Retained=320, Compression Rate=88.9%2026.01 | 0.912 | 91.2 | |
| ViTCoPTokens Retained=160, Compression Rate=94.4%2026.01 | 0.844 | 84.4 | |
| SparseVLMTokens Retained=320, Compression Rate=88.9%2026.01 | 0.839 | 83.9 | |
| VisionZipTokens Retained=320, Compression Rate=88.9%2026.01 | 0.826 | 82.6 | |
| VisionZipTokens Retained=160, Compression Rate=94.4%2026.01 | 0.697 | 69.7 | |
| FastVTokens Retained=320, Compression Rate=88.9%2026.01 | 0.629 | 62.9 | |
| PyramidDropTokens Retained=320, Compression Rate=88.9%2026.01 | 0.625 | 62.5 |