SCULPT: Subtractive Composition for 3D Part Generation

TL;DR

SCULPT starts from a complete 3D object and repeatedly extracts one coherent part while regenerating the remainder, producing an adaptive number of editable parts with consistent shared boundaries.

Abstract

Part-aware 3D generation should create complete, coherent objects while exposing structural components for editing, materials, animation, and reuse. Prior approaches either segment an already generated shape or add separately generated parts, which can fix boundaries too early or introduce gaps, intersections, and appearance discontinuities. SCULPT instead formulates part generation as subtractive composition in a structured 3D latent space. At each step, a joint split predictor generates an extracted part together with the updated remainder, conditioned on the input image and current 3D state. Both outputs are denoised jointly on the union of their sparse supports, allowing neighboring supports to overlap without forcing a disjoint voxel partition. Generation stops when the remainder becomes empty or reaches a safety cap, so part count adapts to the object. Experiments show state-of-the-art geometry on PartObjaverse, strong reconstruction after assembly, and fine-grained textured decomposition on benchmark, generated, and real-world images.

Key contributions

• Introduces subtractive composition for adaptive part-aware 3D generation.
• Jointly predicts each extracted part and its remaining object in structured 3D latent space.
• Maintains high-quality complete-object reconstruction while enabling textured part decomposition.

Citation

@article{li2026sculpt,
  title={SCULPT: Subtractive Composition for 3D Part Generation},
  author={Li, Sikuang and Yang, Chen and Fang, Jiemin and Cen, Jiazhong and Wei, Yuhe and Pang, Jichen and Shen, Wei and Tian, Qi},
  journal={arXiv preprint arXiv:2608.13541},
  year={2026},
  doi={10.48550/arXiv.2608.13541}
}