Reproduce “DetailGen3D: Generative 3D Geometry Enhancement via Data-Dependent Flow” 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
Ken Deng, Yuan-Chen Guo, Jingxiang Sun, Zi-Xin Zou, Yangguang Li, Xin Cai, Yan-Pei Cao, Yebin Liu and 1 more
DetailGen3D, from Tsinghua University and VAST researchers, introduces a generative method to enhance coarse 3D shapes with high-fidelity details by learning a data-dependent rectified flow in a latent space. The approach consistently improved visual quality across various 3D generation and reconstruction inputs, achieving lower FID scores and strong user preference.
Key insights
- Directly modeling the transformation path between coarse and fine geometry in latent space via rectified flow is more efficient than noise-to-data mapping.
- Maintaining precise spatial correspondence through "token matching" is crucial for adding localized details without distorting global shape.
- Generating coarse-fine training pairs using Large Reconstruction Models (LRMs) from high-quality 3D data effectively mimics real-world generation artifacts, enhancing robustness and generalizability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.