Video Understanding on WorldSense (Efficiency & Latency Focused)
12TFLOPsSEATS
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| SEATSR=10, Backbone=Qwen2.5-Omni-7B2026.05 | 12 | 17.65 | 0.019 | 0.062 | 0.196 | 0.767 | 46.9 | |
| SEATSR=25, Backbone=Qwen2.5-Omni-7B2026.05 | 26.5 | 18.29 | 0.028 | 0.082 | 0.342 | 0.923 | 48.6 | |
| SEATSR=35, Backbone=Qwen2.5-Omni-7B2026.05 | 36.7 | 18.68 | 0.034 | 0.092 | 0.436 | 1.023 | 49.3 | |
| VisionZip-omR=35, Backbone=Qwen2.5-Omni-7B2026.05 | 37.4 | 19.63 | 0.125 | — | 0.349 | 1.066 | 48.3 | |
| DivPrune-omR=35, Backbone=Qwen2.5-Omni-7B2026.05 | 37.4 | 19.63 | 0.025 | — | 0.349 | 0.952 | 48.3 | |
| OmniZipR=35, Backbone=Qwen2.5-Omni-7B2026.05 | 38.2 | 19.69 | 0.095 | — | 0.36 | 1.088 | 48.2 | |
| FastV-omR=35, Backbone=Qwen2.5-Omni-7B2026.05 | 38.3 | 19.48 | — | 0.01 | 0.363 | 0.901 | 47.2 | |
| Full tokensR=100, Backbone=Qwen2.5-Omni-7B2026.05 | 111 | 22.83 | — | — | 0.937 | 1.471 | 48.7 |