SCULPT: SUBTRACTIVE COMPOSITION FOR 3D PART GENERATION

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, leading to various challenges in geometry preservation and part consistency. In this paper, we propose SCULPT, a novel framework that addresses these issues through subtractive composition. SCULPT iteratively applies a joint split predictor to generate an extracted part alongside the remaining object from a complete object represented in a structured 3D latent space. The predictor uses a coupled denoising process conditioned on both the image and the current 3D state to collaboratively generate the extracted part and updated remainder. This approach allows overlapping neighboring supports, facilitating a more integrated generation process. The rollout concludes when the remainder support is empty or reaches capacity, which adapts the number of generated parts to each object. Extensive experiments demonstrate that SCULPT achieves state-of-the-art geometry on PartObjects while ensuring robust complete-object reconstruction post-part assembly. Results on various datasets illustrate the framework's ability for fine-grained textured part decomposition, surpassing prior benchmarks.