V-PRUNE

Semantic-aware patch pruning for efficient LVLM inference.

Overview

V-PRUNE removes redundant visual patches before tokenization.

Features

- Training-free
- Patch-level pruning
- FLOPs reduction
- Faster inference

Method

1. Patch grouping
2. Similarity analysis
3. Redundant patch removal
4. Standard LVLM inference

Results

- >25% FLOPs reduction
- >95% accuracy retention

Publication

Applied Sciences

Citation

@article{seo2025v,
  title={V-PRUNE: Semantic-Aware Patch Pruning Before Tokenization in Vision--Language Model Inference},
  author={Seo, Hyein and Choi, Yong Suk},
  journal={Applied Sciences},
  volume={15},
  number={17},
  pages={9463},
  year={2025},
  publisher={MDPI}
}```