Reproduce “Foveated Compression: Selective High-Resolution Preservation for Token-Efficient VLMs” 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
Donghyun Han, Jangho Park, Yuseok Bae
Foveated Compression allocates visual tokens by compressing image features and restoring one selected region to native-resolution tokens. The frozen vision encoder runs once; a learned selector chooses among nine cells. On eight benchmarks, macro accuracy (official scores averaged with MME divided by 28) was 69.61, above random selection’s 67.98 at the same 20.99% token budget, but below whole-image resizing’s 70.99 at 19.82% tokens.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.