Pre-Conference Talk by YU Mingzhe | Dual-Diffusional Generative Fashion Recommendation

Personalized generative recommender systems have emerged as a promising approach to fashion recommendation. However, existing methods primarily rely on implicit visual embeddings derived from historical interactions. These embeddings often contain preference-irrelevant information, leading to insufficient user behavior modeling. Moreover, existing models typically generate only item images, offering limited interpretability. To address these limitations, we propose DualFashion, a dual-diffusional generative fashion recommendation architecture that jointly models image and text modalities for personalized and explainable recommendation. DualFashion employs a dual-diffusion Transformer comprising image and text branches. Structured attribute-level captions and visual outfit information are jointly incorporated as conditioning signals to model user behavior. The architecture generates both fashion item images and textual descriptions, ensuring visual compatibility while providing explicit semantic explanations. We further introduce a text-augmented fine-tuning strategy that improves generation diversity and facilitates effective cross-modal knowledge transfer without substantial computational overhead. Extensive experiments on iFashion and Polyvore-U demonstrate that DualFashion achieves strong performance in user behavior modeling, interpretability, and computational efficiency compared with state-of-the-art methods.

This is a Pre-Conference talk for The 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2026).