Inference Efficiency on POPE (test)
756,000Total Time (ms)VisionZip
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| VisionZipToken=160, Backbone=LLaVA-NeXT 7B, Hardware=single NVIDIA A800-80GB2024.12 | 756,000 | — | 27.8 | — | 3 | 7.8 | |
| FastVToken=160, Backbone=LLaVA-NeXT 7B, Hardware=single NVIDIA A800-80GB2024.12 | 1,792,000 | — | 119 | — | 1.3 | 1.8 | |
| SparseVLMToken=160, Backbone=LLaVA-NeXT 7B, Hardware=single NVIDIA A800-80GB2024.12 | 1,895,000 | — | 128 | — | 1.2 | 1.7 | |
| BaselineToken=2880, Backbone=LLaVA-NeXT 7B, Hardware=single NVIDIA A800-80GB2024.12 | 2,293,000 | — | 218 | — | — | — |