Reproduce “Beyond Selection: Token Parameterization for Extreme Visual Token Compression” 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
Rui Zhong, Yu Li, Zheyu Yan, Cheng Zhuo
Braco compresses image-token grids into a low-frequency DCT backbone, adds coordinate cues, and appends learned spatial residual tokens. On LLaVA-1.5-7B, 16 of 576 tokens produce 94.0 Acc., the mean of eight scores normalized by corresponding uncompressed Vanilla scores, versus QueCC’s 93.9; full-pipeline prefill latency is 40.59 versus 63.22 milliseconds. Evidence primarily concerns image VLM pipelines and falls below QueCC at four tokens.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.