SCULPT: Subtractive Composition
for 3D Part Generation

Sikuang Li1,*, Chen Yang2,*, Jiemin Fang2,✉,§, Jiazhong Cen1,
Yuhe Wei1, Jichen Pang1, Wei Shen1,✉, Qi Tian2,✉
1Shanghai Jiao Tong University, 2Huawei
*Equal contribution. ✉Corresponding authors. §Project lead.
This work is done during an internship at Huawei.

SCULPT generates complete 3D objects and decomposes them into
semantically meaningful, textured parts — part by part,
inside the generator's own latent space.

arXiv Paper Code

Demo

Interactive Results

Drag to orbit, scroll to zoom, and use the Explode slider to pull the generated parts apart.
Every part below was generated by SCULPT — closed, textured, and pre-aligned in the object frame.
Models are compressed for the web (5–20 MB each) and load on demand.

Loading…

More Results

Exploded views of more generated objects. Each clip shows the complete object bursting into its generated parts.

Complex Cases

SCULPT scales to intricate assemblies with up to hundreds of parts — the part count adapts to each object.

Method

Starting from a complete object generated by a pretrained structured-latent backbone, SCULPT applies a joint split predictor that couples an image-conditioned denoising branch with a ControlNet-style branch injecting the current 3D state. At every step it generates one extracted part together with the updated remainder in a single coupled denoising process, so the two outputs share an interface shell on the native sparse support — adjacent parts stay aligned without welding, snapping, or rescaling.

SCULPT pipeline: joint split predictor with decomposition flow transformer blocks.
The joint split predictor performs coupled denoising conditioned on the input image and the current remainder, emitting an extracted part ẑi and the updated remainder zi at every iteration, until the residual support is empty.
Subtractive composition sequence: parts are extracted iteratively until the remainder is empty.
Subtractive composition. Parts (green) are extracted from the remaining object (red) one at a time; the process terminates when the residual sparse support becomes empty, so the part count adapts to each object.

Comparisons

Compared with segmentation-based and additive part-generation baselines (Part123, OmniPart, PartField, SAM3D, HoloPart pipelines), SCULPT preserves global shape while producing coherent part boundaries and appearance, improving part-level Chamfer distance by 7.0% over the strongest baseline on the PartObjaverse benchmark.

Qualitative comparison against part-structured generation baselines.
Qualitative comparison. Insets highlight part decomposition quality at shared boundaries.

BibTeX

@misc{sculpt2026,
  title={SCULPT: Subtractive Composition for 3D Part Generation}, 
  author={Sikuang Li and Chen Yang and Jiemin Fang and Jiazhong Cen and Yuhe Wei and Jichen Pang and Wei Shen and Qi Tian},
  year={2026},
  eprint={2608.13541},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2608.13541}, 
}