Reproduce “Flow-CLIP: Image-Conditioned Residual Scoring for CompositionalImage-Text Matching” 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
Minghong Wei, Enguang Zhu, Pu Cao, Lu Yang, Qing Song
Flow-CLIP addresses image–caption mismatches in which captions name the same objects and attributes but assign them differently. It combines a learned global sentence–image score with image-conditioned corrections to token-level evidence. In matched COCO-trained comparisons, it improves pairwise accuracy on CLEVR’s 449-pair assignment split from 44.99% to 53.67%, but does not improve every spatial task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.