High-fidelity 3D garment reconstruction via dual-detail aware Gaussian splatting and diffusion

3D garment reconstruction is the digital creation of 3D garment models, essential for virtual try-on and digital content creation. Existing 3D garment reconstruction methods often adopt volumetric representations or apply uniform sampling to Gaussian‑splat primitives. These choices under‑sample high‑curvature regions, leading to geometric distortion, misaligned seams, and loss of fine fabric details, while wasting samples on flat areas. We propose high-fidelity 3D garment reconstruction via Dual-Detail Aware Gaussian Splatting and Diffusion, a curvature-aware garment reconstruction framework that holistically addresses the geometric structure and textural appearance of garments. Geometrically, our method dynamically allocates more Gaussian elements to structurally complex regions while preserving efficiency in flat areas, establishing a curvature-driven sampling mechanism that addresses the issue of poor geometric detail recovery caused by uniform sampling. Texturally, to address the prevalent issues of cross-view inconsistency and seam artifacts in existing UV space optimization methods, we introduce a diffusion-based texture refinement mechanism that leverages structural similarity guidance to achieve high-resolution texture generation. Experiments demonstrate that our approach achieves superior performance in geometric accuracy and texture realism compared to state-of-the-art garment reconstruction techniques.