Pipeline
We introduce a core 3D semantic clothing model to enhance controllability in clothing generation, editing, and optimization. (a) Given a text or image prompt, our proposed generation approach leverages the semantic clothing model and an interpolated densification strategy to initialize 3DGS for different clothing types, while leveraging bimodal pretrained 2D diffusion models to refine geometric and texture details. (b) Semantic-based editing can prune and densify the Gaussians according to user requirements, while applying SDS supervision to enhance details. (c) Self-occlusion optimization leverages semantic information to address the self-occlusion problem.