Text-Conditioned Outfit Recommendation With Hybrid Attention Layer

Text-conditioned outfit recommendation aims to recommend a whole fashion outfit that satisfies the compatibility between the recommended items and given items and adheres to the text condition like “Paradise Tropical Vacation” or “60s Style”. Using text description as a condition can provide users with a flexible and accurate way to retrieve and recommend fashion items but this problem is underexplored by existing studies. A challenge of text-conditioned outfit recommendation is how to encode and fuse the outfit text description and fashion item images and text. To solve this, this paper proposes a framework for this task which features a hybrid attention layer that constructs the relationship between outfit text description and fashion items for condition compliance, and the relationship between fashion items for internal compatibility. To encode fashion item features, our method uses pre-trained FashionCLIP as an extractor which significantly reduces the trainable parameters compared to previous methods training CNN from scratch. The whole outfits are generated by iteratively adding compatible items based on a given partial outfit. Compared with state-of-the-art methods on polyvore disjoint and non-disjoint datasets, our approach can achieve 3% relative improvement in compatibility prediction AUC, achieve 5% relative improvement in fill-in-the-blank accuracy; achieve 19% relative improvement on complementary item retrieval recall at different ranks in average. Besides, We demonstrate that our approach can recommend a whole outfit with inner compatibility and adhere to the text description.