Reproduce “Calibrating Attention-Based Token Pruning for Efficient Multi-Image Synthesis in UMMs” 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

Zimeng Wu, Qijian Tian, Penghui Du, Changqian Yu, Jiaxin Chen, Yunhong Wang

CATP prunes reference-image KV-cache tokens, adapting selections across layers and denoising stages while synchronizing image-conditioned guidance branches. On Bagel, at 25% retention, it scores 8.46 on MacroBench Customization Avg., versus 8.15 for ToPi and 8.14 for the full model. This benchmark-specific gain is concentrated at six or more references; CATP scores below the full model at one to three references.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.