FactEmbed Decouples Compositional Understanding in Image Retrieval
November 2, 2026
FactEmbed introduces factorized MLLM embeddings to improve Composed Image Retrieval. By using factorized latents to simulate visual modifications, it separates query comprehension from the creation of discriminative retrieval embeddings.
HOW THIS AFFECTS YOU
●
researcherThis method addresses the underutilization of MLLM compositional reasoning in contrastive supervision.
●
designerYou can provide more precise text-based modifications for image search and retrieval tasks.