SCULPT: Subtractive Composition for Part-Aware 3D Generation

Forum topic · 小凯 · 2026-08-15

Summary

SCULPT is a framework for part-aware 3D generation that produces digital assets coherent as complete objects while exposing structural parts for editing, material assignment, animation, and reuse. Existing approaches impose part structure outside the native generation loop—either segmenting an already generated shape or additively composing parts from predefined layouts, boxes, or tokens and then reconciling them. These methods either freeze objects before part boundaries are set or leave shared boundaries prone to gaps, interpenetrations, and material discontinuities. SCULPT instead uses subtractive composition: given a complete object in a structured 3D latent space, it iteratively applies a joint split predictor that produces one extracted part together with the updated remainder. The predictor performs coupled denoising conditioned on the image and current 3D state, so both outputs are generated jointly. It operates on the union of native sparse 3D supports, allowing neighboring supports to overlap rather than enforcing disjoint voxel partitions. Rollout stops when the remainder is empty or reaches a safety cap, letting part count adapt per object.

Paper Overview

Research Area: Computer Vision (CV)
Authors: Sikuang Li, Chen Yang, Jiemin Fang, Jiazhong Cen, Yuhe Wei, Jichen Pang, Wei Shen, Qi Tian
Release Date: 2026-08-13
arXiv: 2608.13541

Abstract

Part-aware 3D generation aims to create digital assets that are coherent as complete objects while exposing structural parts for editing, material assignment, animation, and reuse. Existing methods impose this structure outside the native generation loop: segmentation-based methods partition an already generated shape, while additive methods synthesize parts from predefined layouts, boxes, or tokens and then reconcile them into a whole. The former preserves the generated geometry but fixes the object before part boundaries are determined; the latter exposes part cardinality but often leaves shared boundaries vulnerable to gaps, interpenetrations, and material discontinuities.

In this paper, we propose SCULPT, a framework that addresses these challenges through subtractive composition. Given a complete object represented in a structured 3D latent space, SCULPT iteratively applies a joint split predictor to generate one extracted part together with the remaining object. The predictor performs a coupled denoising process conditioned on both the image and the current 3D state, so the extracted part and updated remainder are generated together rather than reconciled after generation. The joint split predictor processes both outputs on the union of their native sparse 3D supports, allowing neighboring supports to overlap rather than imposing a disjoint voxel partition. The rollout ends when the remainder support becomes empty or reaches a fixed safety cap, allowing the number of generated parts to adapt to each object within that bound.

Key Contributions

Subtractive composition paradigm: Shifts from post-hoc segmentation or additive assembly to iterative, in-the-loop part extraction directly within the 3D generation process.
Joint split predictor: Performs coupled denoising conditioned on both the input image and the current 3D state, generating the extracted part and the updated remainder simultaneously.
Overlapping sparse 3D support: Processes both outputs on the union of their native sparse supports, permitting neighboring parts to overlap and avoiding hard disjoint voxel partitions.
Adaptive part cardinality: Stops iterative splitting when the remainder support becomes empty or hits a fixed safety cap, so the part count varies per object.

Experimental Highlights

State-of-the-art geometry on the PartObjaverse dataset, while preserving strong complete-object reconstruction after part assembly.
Generalization across input types: results on four dataset images, one text-to-image-generated input, and one real-world photograph.
Fine-grained textured part decomposition that goes beyond the benchmark setting.

Method Notes

Operates in a structured 3D latent space representing the full object.
Uses coupled denoising so the part and remainder are coherent at generation time, not stitched post hoc.
Avoids common artifacts (gaps, interpenetration, material discontinuities) by allowing supports to overlap instead of forcing voxel-disjoint boundaries.