Video Compression Commander
A plug-and-play token-compression framework for video LLMs that dynamically allocates a frame token budget, cutting a model such as LLaVA-OV's generation time by over 70 percent while keeping most accuracy, unifying several published baselines.
Video Compression Commander, or VidCom2, is a research framework for accelerating inference in video large language models by compressing the number of visual tokens a model has to process, using dynamic frame budget allocation rather than a fixed compression rate per frame. It is plug-and-play: it drops into existing video LLMs without retraining.
It reports using only about 25 percent of the original tokens while retaining 99.6 percent of LLaVA-OV's benchmark performance, cutting LLaVA-OV's generation time by 70.8 percent and its overall latency by 43.0 percent, and it is compatible with efficient attention operators such as Flash Attention 2. The codebase also integrates several other published token-compression baselines, including VisionZip, DyCoke, FastVID and HoliTom, into one unified framework, with support for the LLaVA, Qwen-VL and Qwen-Omni model families.
It accompanies a paper accepted at the EMNLP 2025 conference, and the same team has released related follow-up work on curvature-aware pruning and streaming video understanding.