Production-ready texture synthesis from a single image

Establish a method for generating production-ready textures for 3D garment assets from a single observation, with textures that preserve garment appearance while avoiding baked-in illumination artifacts, geometric distortion, and loss of global structural coherence.

Background

The paper identifies high-quality texture generation as a major remaining obstacle in automatically reconstructing simulation-ready 3D garments from single images. Existing approaches commonly project textures onto draped 3D geometry, which can introduce distortion, occlusion artifacts, shadows, and other illumination effects into the resulting texture maps.

The stated challenge is particularly significant for garment assets because production workflows require globally coherent textures that remain suitable for physical simulation, relighting, and downstream digital-content applications. The paper proposes OmniFabric as an approach to this broader unresolved problem by synthesizing textures in the 2D sewing-pattern UV domain.