Reproduce “Factorized MLLM Embeddings for Composed Image Retrieval” 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
WonJun Moon, Seongchan Kim, Dahun Kim, Seungryong Kim
FactEmbed addresses composed image retrieval by adding eight latent factors that reconstruct target features before search. A decoder uses a coarse 2×2 layout prior, while factor-level contrastive training aligns query and target factors. On Refined-CIRR, it reports 68.99% R@1 versus ReCALL’s 67.41%, and a Refined-FashionIQ average Recall@10 of 71.13 versus 69.48. These results include augmented post-training and apply to refined evaluations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.