Reproduce “Joint Geometry-Appearance Human Reconstruction in a Unified Latent Space via Bridge Diffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yingzhi Tang, Qijian Zhang, Junhui Hou
Researchers at City University of Hong Kong and Tencent Games developed JGA-LBD, a framework that unifies 3D human geometry and appearance reconstruction from a single RGB image into a joint latent space using bridge diffusion. This approach achieved state-of-the-art results, with a PSNR of 30.16 and SSIM of 0.946 for appearance, and a Chamfer Distance of 0.489 for geometry on the 2K2K benchmark.
Key insights
- Unifying geometry and appearance in a single, compact latent space for 3D Gaussians inherently ensures consistency and enables a coherent, single-stage reconstruction process.
- Bridge diffusion effectively leverages partial observations (e.g., depth-derived latent codes) to guide the generation process, focusing on inferring missing information and reducing generative complexity.
- Transforming heterogeneous 3D structural priors into a consistent 3D Gaussian format is critical for stabilizing training and enabling effective conditioning of latent diffusion models for complex 3D tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.