V-CAST: training-free video token pruning keeps 97% accuracy with 51% faster prefill
量子位 · wechat · 2026-11-17
Researchers from Sun Yat-sen University, PolyU and others propose V-CAST, a training-free, plug-and-play spatio-temporal pruning method for video LLMs.
Design:
- Allocates per-frame token budgets via local curvature of frame features, giving more tokens to moments of action/scene change.
- Selects in-frame tokens with two complementary scores: deviation from frame mean (keeps small local details) and activation strength.
- Keeps original token order and spatio-temporal coordinates; direct selection beats local pooling.
Results across Qwen3-VL, LLaVA and Qwen2.5-Omni (6 configs, 10 benchmarks):
- 25% token retention keeps 98.2% of Qwen3-VL-8B performance; leads next-best by 1.5 pts at 256 frames.
- 768-frame audio-video experiment retains 97.1% accuracy with 50% lower total inference time.
- Prefill latency drops 51.2%, peak memory 13.0%; enables 64-frame eval that OOMs uncompressed.
- The curvature budgeting also boosts VisionZip and VidCom² when transplanted.
Code and project page are open-sourced.
More from Multimodal
- Recreating The Godfather's office scene with MiniMax H3 in ComfyUI: 3 shots, 37s — hucun16 · 2026-11-17
- Creator turns a Magnific Spaces retro-video workflow into an unhinged 1973 canned-bean commercial — charis_ai · 2026-11-17
- HexAnima fork runs offline image generation on Snapdragon 8 Gen 2 NPU at ~7s/image — ilinkys · 2026-11-17
- AI-generated black-and-red colossus face hides a burning city in its eye socket — LudovicCreator · 2026-11-17
- AI-Animated Series "The Croaky Inn" Returns With Stunning Toad Character in Episode 2 — jollypiraterum · 2026-11-17
- Idea renders Hollywood-grade 3D animation from your AI chat for $15/month — wordgrammer · 2026-11-17