Reproduce “SymTRELLIS: Symmetry-Enforced Voxel Latents for 3D Generation” 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
Guangda Ji, Qimin Chen, Qinchan Li, Mingrui Zhao, Kai Wang, Hao Zhang
SymTRELLIS integrates explicit finite-group symmetry into TRELLIS.2’s image-to-3D generation process by learning how rotations and reflections act on voxel latents, rather than repairing the final mesh afterward. On 266 symmetric objects, it reduced average symmetry error from 4.58 to 2.24 × 10⁻³, while automatic symmetry detection reached 83.81% fold accuracy and added under 0.15 seconds of mapper inference on an RTX 5090.
Key insights
- Enforcing symmetry during generation is more effective than averaging the completed mesh because image-conditioned flow updates can reconstruct and coordinate unseen sectors.
- Sparse-structure guidance provides most of the global symmetry improvement, while additional shape-stage guidance improves local consistency but may smooth high-frequency details.
- A learned latent transformation operator makes group-based constraints possible even when the internal feature channels of a generative model have unknown geometric behavior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.