Reproduce “Hybrid Token Compression for Vision-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jusheng Zhang, Xiaoyang Guo, Tongyu Mo, Qinhan Lv, Wenhao Chai, Jian Wang, Keze Wang, Liang Lin
The authors introduce HTC-VLM, compressing image inputs to one visual token while preserving both semantic cues and appearance details. It combines four MGVQ-derived anchors with 576 continuous patch features, then routes them through a shared bottleneck. In LLaVA-1.5, it achieves 87.2% average normalized performance retention, versus 81.0% for continuous-only compression; this advantage is reported for the one-token setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.