Multi-Granular Spatio-Temporal Token Merging for Training-Free Acceleration of Video LLMs

TL;DR

The paper introduces STTM, a training-free spatio-temporal token merging method that exploits local redundancies in video data to accelerate Video LLMs without significant accuracy loss.

Contribution

STTM is a novel, training-free token merging approach that improves video understanding efficiency by leveraging local redundancies and multi-granular spatial tokens.

Findings

Achieves 2x speed-up with 0.5% accuracy drop at 50% token budget.

Achieves 3x speed-up with 2% accuracy drop at 30% token budget.

Outperforms existing token reduction methods across six video QA benchmarks.

Abstract

Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2$\times$ speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3$\times$ speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The project page is available at https://www.jshyun.me/projects/sttm.