The paper introduces a novel task called Generative Outfit Recommendation (GOR), which aims to generate a set of fashion images and compose them into a visually compatible and personalized outfit tailored to specific users. To achieve this, the authors propose DiFashion, a generative outfit recommender model that leverages diffusion models to generate multiple fashion images in parallel. The key objectives of GOR are high fidelity, compatibility, and personalization. DiFashion employs three conditions—category prompt, mutual condition, and history condition—to guide the parallel generation process, ensuring that the generated outfits align with user preferences and are visually compatible. Extensive experiments on the iFashion and Polyvore-U datasets demonstrate the superiority of DiFashion over competitive baselines in both personalized Fill-In-The-Blank (PFITB) and GOR tasks. The paper also includes a human-involved qualitative evaluation and an in-depth analysis of the impact of various design choices in DiFashion. The results show that DiFashion consistently outperforms baselines in meeting the three criteria of GOR. The work contributes to a more personalized fashion landscape by generating new, tailored fashion products that can be retrieved or customized for practical use.The paper introduces a novel task called Generative Outfit Recommendation (GOR), which aims to generate a set of fashion images and compose them into a visually compatible and personalized outfit tailored to specific users. To achieve this, the authors propose DiFashion, a generative outfit recommender model that leverages diffusion models to generate multiple fashion images in parallel. The key objectives of GOR are high fidelity, compatibility, and personalization. DiFashion employs three conditions—category prompt, mutual condition, and history condition—to guide the parallel generation process, ensuring that the generated outfits align with user preferences and are visually compatible. Extensive experiments on the iFashion and Polyvore-U datasets demonstrate the superiority of DiFashion over competitive baselines in both personalized Fill-In-The-Blank (PFITB) and GOR tasks. The paper also includes a human-involved qualitative evaluation and an in-depth analysis of the impact of various design choices in DiFashion. The results show that DiFashion consistently outperforms baselines in meeting the three criteria of GOR. The work contributes to a more personalized fashion landscape by generating new, tailored fashion products that can be retrieved or customized for practical use.