FastVID: Dynamic Density Pruning for Fast Video Large Language Models
This work addresses deployment challenges for video AI systems by reducing computational overhead, though it is incremental as it builds on existing pruning techniques.
The paper tackles the high inference costs of Video Large Language Models by addressing redundant video tokens, proposing FastVID, which dynamically prunes tokens to reduce FLOPs to 8.3% and accelerate prefilling by 7.1x while maintaining 98.0% accuracy on LLaVA-OneVision-7B.
Video Large Language Models have demonstrated strong video understanding capabilities, yet their practical deployment is hindered by substantial inference costs caused by redundant video tokens. Existing pruning techniques fail to fully exploit the spatiotemporal redundancy inherent in video data. To bridge this gap, we perform a systematic analysis of video redundancy from two perspectives: temporal context and visual context. Leveraging these insights, we propose Dynamic Density Pruning for Fast Video LLMs termed FastVID. Specifically, FastVID dynamically partitions videos into temporally ordered segments to preserve temporal structure and applies a density-based token pruning strategy to maintain essential visual information. Our method significantly reduces computational overhead while maintaining temporal and visual integrity. Extensive evaluations show that FastVID achieves state-of-the-art performance across various short- and long-video benchmarks on leading Video LLMs, including LLaVA-OneVision and LLaVA-Video. Notably, on LLaVA-OneVision-7B, FastVID effectively prunes $\textbf{90.3%}$ of video tokens, reduces FLOPs to $\textbf{8.3%}$, and accelerates the prefilling stage by $\textbf{7.1}\times$, while maintaining $\textbf{98.0%}$ of the original accuracy. The code is available at https://github.com/LunarShen/FastVID.