IPCV: Information-Preserving Compression for MLLM Visual Encoders

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Chen, Yuan, Wen, Zichen, Wu, Yuzhou, Liu, Xuyang, Chen, Shuang, Ma, Junpeng, Li, Weijia, He, Conghui, Zhang, Linfeng
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908726095511552
author Chen, Yuan
Wen, Zichen
Wu, Yuzhou
Liu, Xuyang
Chen, Shuang
Ma, Junpeng
Li, Weijia
He, Conghui
Zhang, Linfeng
author_facet Chen, Yuan
Wen, Zichen
Wu, Yuzhou
Liu, Xuyang
Chen, Shuang
Ma, Junpeng
Li, Weijia
He, Conghui
Zhang, Linfeng
contents Multimodal Large Language Models (MLLMs) deliver strong vision-language performance but at high computational cost, driven by numerous visual tokens processed by the Vision Transformer (ViT) encoder. Existing token pruning strategies are inadequate: LLM-stage token pruning overlooks the ViT's overhead, while conventional ViT token pruning, without language guidance, risks discarding textually critical visual cues and introduces feature distortions amplified by the ViT's bidirectional attention. To meet these challenges, we propose IPCV, a training-free, information-preserving compression framework for MLLM visual encoders. IPCV enables aggressive token pruning inside the ViT via Neighbor-Guided Reconstruction (NGR) that temporarily reconstructs pruned tokens to participate in attention with minimal overhead, then fully restores them before passing to the LLM. Besides, we introduce Attention Stabilization (AS) to further alleviate the negative influence from token pruning by approximating the K/V of pruned tokens. It can be directly applied to previous LLM-side token pruning methods to enhance their performance. Extensive experiments show that IPCV substantially reduces end-to-end computation and outperforms state-of-the-art training-free token compression methods across diverse image and video benchmarks. Our code is available at https://github.com/Perkzi/IPCV.
format Preprint
id arxiv_https___arxiv_org_abs_2512_18747
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle IPCV: Information-Preserving Compression for MLLM Visual Encoders
Chen, Yuan
Wen, Zichen
Wu, Yuzhou
Liu, Xuyang
Chen, Shuang
Ma, Junpeng
Li, Weijia
He, Conghui
Zhang, Linfeng
Computer Vision and Pattern Recognition
Artificial Intelligence
Multimodal Large Language Models (MLLMs) deliver strong vision-language performance but at high computational cost, driven by numerous visual tokens processed by the Vision Transformer (ViT) encoder. Existing token pruning strategies are inadequate: LLM-stage token pruning overlooks the ViT's overhead, while conventional ViT token pruning, without language guidance, risks discarding textually critical visual cues and introduces feature distortions amplified by the ViT's bidirectional attention. To meet these challenges, we propose IPCV, a training-free, information-preserving compression framework for MLLM visual encoders. IPCV enables aggressive token pruning inside the ViT via Neighbor-Guided Reconstruction (NGR) that temporarily reconstructs pruned tokens to participate in attention with minimal overhead, then fully restores them before passing to the LLM. Besides, we introduce Attention Stabilization (AS) to further alleviate the negative influence from token pruning by approximating the K/V of pruned tokens. It can be directly applied to previous LLM-side token pruning methods to enhance their performance. Extensive experiments show that IPCV substantially reduces end-to-end computation and outperforms state-of-the-art training-free token compression methods across diverse image and video benchmarks. Our code is available at https://github.com/Perkzi/IPCV.
title IPCV: Information-Preserving Compression for MLLM Visual Encoders
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2512.18747