TextureDiff: 3D-Agnostic Garment Rendering with UV-Aligned Texture Diffusion

Dressed-human rendering has become a focal research topic due to the growing accessibility and scalability of e-commerce content-generation pipelines. Existing methods often depend on precise 3D reconstruction or UV maps. Consequently, their performance is highly sensitive to the quality of these prerequisites. In this paper, we propose a diffusion-based texture-rendering model, TextureDiff. TextureDiff synthesizes photorealistic images and generalizes well to unseen test cases. To address misalignment between texture patches and the target image structure, TextureDiff first uses UV rendering to convert textures into semantically aligned features. It then treats texture-derived features as prompts, avoiding information loss introduced by textual conversion and hand-crafted alignment. Additionally, we condition the model on Canny edge maps of the target image, which helps preserve fine garment wrinkles and structure. Experiments show that TextureDiff outperforms state-of-theart diffusion-based image-prompting baselines on both real and synthetic datasets.