HieraVid: Hierarchical Token Pruning for Fast Video Large Language Models

TL;DR

HieraVid introduces a hierarchical token pruning method that significantly reduces computational load in Video Large Language Models by exploiting video structure, achieving high performance with only 30% of tokens.

Contribution

The paper proposes a novel hierarchical pruning framework that dynamically reduces visual redundancy in VideoLLMs by leveraging video structure and LLM information propagation.

Findings

Achieves state-of-the-art performance with only 30% tokens retained.

Maintains over 98% and 99% of LLaVA-Video-7B and LLaVA-OneVision-7B performance.

Effective across four widely used video understanding benchmarks.

Abstract

Video Large Language Models (VideoLLMs) have demonstrated impressive capabilities in video understanding, yet the massive number of input video tokens incurs a significant computational burden for deployment. Existing methods mainly prune video tokens at input level while neglecting the inherent information structure embedded in videos and large language models (LLMs). To address this, we propose HieraVid, a hierarchical pruning framework that progressively and dynamically reduces visual redundancy. Based on two observations that videos possess the segment-frame structure and LLMs internally propagate multi-modal information unidirectionally, we decompose pruning into three levels: 1) segment-level, where video tokens are first temporally segmented and spatially merged; 2) frame-level, where similar frames within the same segment are jointly pruned to preserve diversity; 3) layer-level, redundancy gradually shrinks as LLM layer increases w/o compromising performance. We conduct extensive experiments on four widely used video understanding benchmarks to comprehensively evaluate the effectiveness of HieraVid. Remarkably, with only 30% of tokens retained, HieraVid achieves new state-of-the-art performance, while maintaining over 98% and 99% of the performance of LLaVA-Video-7B and LLaVA-OneVision-7B, respectively.