Reproduce “SCULPT: Subtractive Composition for 3D Part 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
Sikuang Li, Chen Yang, Jiemin Fang, Jiazhong Cen, Yuhe Wei, Jichen Pang, Wei Shen, Qi Tian
SCULPT introduces a subtractive composition framework for part-aware 3D generation, producing coherent objects with structured, manipulable parts by iteratively extracting them from a complete object within the generative loop. The framework achieved state-of-the-art quantitative geometry comparison on the PartObjaverse benchmark, reducing part-level Chamfer distance from 0.0136 to 0.0107 compared to direct part generation baselines.
Key insights
- Treating part decomposition as a subtractive generative problem, rather than solely a recognition task or an additive assembly, effectively handles variable part cardinality and promotes boundary coherence.
- Jointly predicting both the extracted part and the remaining object during the generative process allows for the determination and refinement of part interfaces, avoiding common issues like gaps or material discontinuities.
- Leveraging and adapting priors from high-quality holistic 3D generators enables SCULPT to integrate complex object structures and appearance into the decomposition process while maintaining overall object coherence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.