Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Li, Yulin, Gui, Haokun, Fan, Ziyang, Wang, Junjie, Kang, Bin, Chen, Bin, Tian, Zhuotao
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866912753424269312
author Li, Yulin
Gui, Haokun
Fan, Ziyang
Wang, Junjie
Kang, Bin
Chen, Bin
Tian, Zhuotao
author_facet Li, Yulin
Gui, Haokun
Fan, Ziyang
Wang, Junjie
Kang, Bin
Chen, Bin
Tian, Zhuotao
contents Recent advances in Video Large Language Models (VLLMs) have achieved remarkable video understanding capabilities, yet face critical efficiency bottlenecks due to quadratic computational growth with lengthy visual token sequences of long videos. While existing keyframe sampling methods can improve temporal modeling efficiency, additional computational cost is introduced before feature encoding, and the binary frame selection paradigm is found suboptimal. Therefore, in this work, we propose Dynamic Token compression via LLM-guided Keyframe prior (DyToK), a training-free paradigm that enables dynamic token compression by harnessing VLLMs' inherent attention mechanisms. Our analysis reveals that VLLM attention layers naturally encoding query-conditioned keyframe priors, by which DyToK dynamically adjusts per-frame token retention ratios, prioritizing semantically rich frames while suppressing redundancies. Extensive experiments demonstrate that DyToK achieves state-of-the-art efficiency-accuracy tradeoffs. DyToK shows plug-and-play compatibility with existing compression methods, such as VisionZip and FastV, attaining 4.3x faster inference while preserving accuracy across multiple VLLMs, such as LLaVA-OneVision and Qwen2.5-VL. Code is available at https://github.com/yu-lin-li/DyToK .
format Preprint
id arxiv_https___arxiv_org_abs_2512_06866
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior
Li, Yulin
Gui, Haokun
Fan, Ziyang
Wang, Junjie
Kang, Bin
Chen, Bin
Tian, Zhuotao
Computer Vision and Pattern Recognition
Artificial Intelligence
Computation and Language
Machine Learning
Recent advances in Video Large Language Models (VLLMs) have achieved remarkable video understanding capabilities, yet face critical efficiency bottlenecks due to quadratic computational growth with lengthy visual token sequences of long videos. While existing keyframe sampling methods can improve temporal modeling efficiency, additional computational cost is introduced before feature encoding, and the binary frame selection paradigm is found suboptimal. Therefore, in this work, we propose Dynamic Token compression via LLM-guided Keyframe prior (DyToK), a training-free paradigm that enables dynamic token compression by harnessing VLLMs' inherent attention mechanisms. Our analysis reveals that VLLM attention layers naturally encoding query-conditioned keyframe priors, by which DyToK dynamically adjusts per-frame token retention ratios, prioritizing semantically rich frames while suppressing redundancies. Extensive experiments demonstrate that DyToK achieves state-of-the-art efficiency-accuracy tradeoffs. DyToK shows plug-and-play compatibility with existing compression methods, such as VisionZip and FastV, attaining 4.3x faster inference while preserving accuracy across multiple VLLMs, such as LLaVA-OneVision and Qwen2.5-VL. Code is available at https://github.com/yu-lin-li/DyToK .
title Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior
topic Computer Vision and Pattern Recognition
Artificial Intelligence
Computation and Language
Machine Learning
url https://arxiv.org/abs/2512.06866