FrameFusion: Combining Similarity and Importance for Video Token Reduction on Large Vision Language Models
Authors:
- 19 October 2025
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 22654-22663
Abstract
The increasing demand to process long and high-resolution videos significantly burdens Large Vision-Language Models (LVLMs) due to the enormous number of visual tokens. Existing token reduction methods primarily prune tokens based on importance metrics, such as cumulative attention scores. However, even important tokens may exhibit high redundancy caused by similarity among adjacent video frames and repetitive visual elements. To address this limitation, we propose FrameFusion, a novel token reduction approach integrating similarity-based merging with importance-based pruning. We conduct a thorough study on token similarity characteristics, revealing three key insights: (1) spatially corresponding visual tokens between adjacent frames have higher cosine similarities compared to other token pairs; (2) high token similarities prominently decrease in deeper model layers; and (3) token similarity rankings are highly consistent across different layers. Guided by these observations, FrameFusion computes token similarities exclusively between corresponding visual tokens from adjacent frames, applies token merging at initial successive layers followed by pruning in deeper layers, and adopts a cascaded merging strategy to further enhance efficiency. We evaluate FrameFusion comprehensively across six diverse LVLMs, ranging from $2 B$ to $72 B$ parameters, using five video benchmarks encompassing video retrieval, question-answering, and spatial-temporal understanding tasks. Experiments show that FrameFusion reduces visual tokens by 70%, achieving $1.6-3.6 \times$ end-toend speedups, with an average performance impact of less than 3%. Our code is available at https://github.com/thu-nics/FrameFusion.
Keywords
All Related Versions
Funding Information
- National Natural Science Foundation of China (62325405,62104128,U19B2019,U21B2031,61832007,62204164,92364201)
This publication has 18 references indexed in Scilit:
- Efficient multimodal large language models: a surveyVisual Intelligence, 2025
- An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models2024
- FlexAttention for Efficient High-Resolution Vision-Language Models2024
- LongVLM: Efficient Long Video Understanding via Large Language Models2024
- LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models2024
- MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding2024
- MVBench: A Comprehensive Multi-modal Video Understanding Benchmark2024
- SparseViT: Revisiting Activation Sparsity for Efficient High-Resolution Vision Transformer2023
- NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions2021
- Spearman Rank Correlation2005