Reproduce “UniPart: Part-Level 3D Generation with Unified 3D Geom-Seg Latents” 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
Xufan He, Yushuang Wu, Xiaoyang Guo, Chongjie Ye, Jiaqing Zhou, Tianlei Hu, Xiaoguang Han, Dong Du
UniPart is a two-stage latent diffusion framework that generates controllable, semantically structured 3D shapes with parts from image prompts. It unifies geometry and part segmentation into a single "Geom-Seg VecSet" latent representation, achieving superior quantitative performance across geometry and segmentation metrics compared to existing part-level 3D generation methods.
Key insights
- Part awareness can emerge implicitly during the learning process of whole-object geometry, suggesting that direct integration of segmentation into the latent space is feasible without requiring explicit segmenters.
- Unifying geometry and segmentation within a single latent representation allows for inherent part reasoning and conditioning in the generative process, preserving reconstruction quality while enabling explicit part control.
- Generating detailed part geometries in a normalized canonical space and then transforming them into global coordinates effectively mitigates fidelity loss for small or intricate parts and ensures consistent assembly.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.