Factorized MLLM Embeddings for Composed Image Retrieval
FactEmbed trains an image-search model to resolve the requested visual changes before building its retrieval representation.
The system searches for a target image using a reference image and text describing a modification. Its intermediate “target imagination” objective trains separate latent representations to capture the intended result. A single model pass produces both a global embedding for retrieval and factorized representations for reranking a shortlist, which the authors say adds negligible cost. They report outperforming prior methods across composed image retrieval benchmarks, though the abstract gives no numerical results.
HF Daily Papers' note
The system searches for a target image using a reference image and text describing a modification. Its intermediate “target imagination” objective trains separate latent representations to capture the intended result. A single model pass produces both a global embedding for retrieval and factorized representations for reranking a shortlist, which the authors say adds negligible cost. They report outperforming prior methods across composed image retrieval benchmarks, though the abstract gives no numerical results.
HF Daily Papers' note
score 4