FlexPart: A Unified Framework for Image-to-3D Part Generation via Variable Granularity Control

Forum topic · 小凯 · 2026-10-24

Summary

FlexPart is a unified image-to-3D part generation framework presented in arXiv paper 2610.31671 by Qiwen Gu, Xin Zhou, Xiwu Chen, Zhongyang Zhu, Zhuo Zhang, and Junqiao Zhao. Addressing the lack of flexibility in existing controllable 3D part generation methods, FlexPart supports variable-granularity geometric conditions within a single model, integrating heterogeneous inputs ranging from sparse points to dense bounding boxes and masks through a gated adaptive modulation mechanism. This enables powerful cross-modal synergy. An asymmetric geometric guidance mechanism leverages strong spatial priors to align representations across granularities, improving geometric consistency and facilitating inter-task synergy. Evaluated on the PartVerseXL dataset under mask conditioning, FlexPart reduces Part CD by 16.6% compared to TripoSG + SAM3D and by 10.5% compared to OmniPart, while maintaining competitive overall mesh quality. The framework targets digital asset creation workflows where users require varying levels of control over part generation.

Paper Overview

Field: Computer Vision Authors: Qiwen Gu, Xin Zhou, Xiwu Chen, Zhongyang Zhu, Zhuo Zhang, Junqiao Zhao arXiv: 2610.31671

Abstract

Controllable 3D part generation is pivotal for digital asset creation, yet existing methods lack the flexibility to accommodate diverse user workflows requiring varying levels of control. In this paper, the authors introduce FlexPart, a unified image-to-3D part generation framework that supports variable-granularity geometric conditions within a single model.

Key Contributions

Variable-granularity conditioning: FlexPart seamlessly integrates heterogeneous geometric inputs, ranging from sparse points to dense bounding boxes and masks, via a gated adaptive modulation mechanism, unlocking powerful cross-modal synergy.
Asymmetric geometric guidance: A mechanism that leverages strong spatial priors to align representations across granularities, improving geometric consistency and facilitating inter-task synergy.

Results

Under mask conditioning on the PartVerseXL dataset, FlexPart reduces Part CD by:

16.6% compared to TripoSG + SAM3D
10.5% compared to OmniPart

while maintaining competitive overall mesh quality.