InfiniteVL: Synergizing Linear and Sparse Attention for Highly-Efficient, Unlimited-Input Vision-Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Tao, Hongyuan, Liao, Bencheng, Chen, Shaoyu, Yin, Haoran, Zhang, Qian, Liu, Wenyu, Wang, Xinggang
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915902281220096
author Tao, Hongyuan
Liao, Bencheng
Chen, Shaoyu
Yin, Haoran
Zhang, Qian
Liu, Wenyu
Wang, Xinggang
author_facet Tao, Hongyuan
Liao, Bencheng
Chen, Shaoyu
Yin, Haoran
Zhang, Qian
Liu, Wenyu
Wang, Xinggang
contents Vision-Language Models (VLMs) are increasingly tasked with ultra-long multimodal understanding. While linear architectures offer constant computation and memory footprints, they often struggle with high-frequency visual perception compared to standard Transformers. To bridge this gap, we introduce \textbf{InfiniteVL}. We first develop a hybrid base model called \textbf{InfiniteVL-Base} that interleaves a small fraction of Full Attention layers with Gated DeltaNet. Empowered by a tailored distillation and fine-tuning strategy, InfiniteVL-Base matches the fundamental multimodal performance of equivalent Transformers while achieving a \textbf{1.7$\times$} decoding speedup. However, the quadratic complexity of the retained Full Attention inevitably becomes an efficiency bottleneck when scaling to ultra long context. To break this barrier, we propose a novel Long-Sequence Architectural Fine-Tuning strategy that seamlessly transforms the dense attention into vision-specific sparse mechanisms. This yields two specialized variants: \textbf{InfiniteVL-Offline} for offline retrieval and \textbf{InfiniteVL-Online} for online streaming. By eliminating the computation explosion of global attention without sacrificing high-frequency visual recall, InfiniteVL-Offline achieves Transformer-level length generalization with a \textbf{5x} prefill acceleration at 256K context. Concurrently, InfiniteVL-Online delivers robust streaming perception with a constant memory footprint and a real-time throughput of \textbf{25} FPS. Code and models are available at https://github.com/hustvl/InfiniteVL.
format Preprint
id arxiv_https___arxiv_org_abs_2512_08829
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle InfiniteVL: Synergizing Linear and Sparse Attention for Highly-Efficient, Unlimited-Input Vision-Language Models
Tao, Hongyuan
Liao, Bencheng
Chen, Shaoyu
Yin, Haoran
Zhang, Qian
Liu, Wenyu
Wang, Xinggang
Computer Vision and Pattern Recognition
Artificial Intelligence
Vision-Language Models (VLMs) are increasingly tasked with ultra-long multimodal understanding. While linear architectures offer constant computation and memory footprints, they often struggle with high-frequency visual perception compared to standard Transformers. To bridge this gap, we introduce \textbf{InfiniteVL}. We first develop a hybrid base model called \textbf{InfiniteVL-Base} that interleaves a small fraction of Full Attention layers with Gated DeltaNet. Empowered by a tailored distillation and fine-tuning strategy, InfiniteVL-Base matches the fundamental multimodal performance of equivalent Transformers while achieving a \textbf{1.7$\times$} decoding speedup. However, the quadratic complexity of the retained Full Attention inevitably becomes an efficiency bottleneck when scaling to ultra long context. To break this barrier, we propose a novel Long-Sequence Architectural Fine-Tuning strategy that seamlessly transforms the dense attention into vision-specific sparse mechanisms. This yields two specialized variants: \textbf{InfiniteVL-Offline} for offline retrieval and \textbf{InfiniteVL-Online} for online streaming. By eliminating the computation explosion of global attention without sacrificing high-frequency visual recall, InfiniteVL-Offline achieves Transformer-level length generalization with a \textbf{5x} prefill acceleration at 256K context. Concurrently, InfiniteVL-Online delivers robust streaming perception with a constant memory footprint and a real-time throughput of \textbf{25} FPS. Code and models are available at https://github.com/hustvl/InfiniteVL.
title InfiniteVL: Synergizing Linear and Sparse Attention for Highly-Efficient, Unlimited-Input Vision-Language Models
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2512.08829