PartSDF: Part-Based Implicit Neural Representation for Composite 3D Shape Parametrization and Optimization

Nicolas Talabot¹, Olivier Clerc¹, Arda Cinar Demirtas², Hieu Le^1,3, Doruk Oner², Pascal Fua¹

¹ CVLab, EPFL

² Bilkent University

³ University of North Carolina at Charlotte

TMLR 2025

Abstract

Accurate 3D shape representation is essential in engineering applications such as design, optimization, and simulation. In practice, engineering workflows require structured, part-based representations, as objects are inherently designed as assemblies of distinct components. However, most existing methods either model shapes holistically or decompose them without predefined part structures, limiting their applicability in real-world design tasks. We propose PartSDF, a supervised implicit representation framework that explicitly models composite shapes with independent, controllable parts while maintaining shape consistency. Thanks to its simple but innovative architecture, PartSDF outperforms both supervised and unsupervised baselines in reconstruction and generation tasks. We further demonstrate its effectiveness as a structured shape prior for engineering applications, enabling precise control over individual components while preserving overall coherence.

Method

(Left) Overview of PartSDF cross-part auto-decoder. (Right) Single- and cross-part layers arrangement in the decoder f_θ.

At its core, our model is a cross-part auto-decoder f_θ that takes as input part latents z_p and poses expressed in terms of a quaternion q_p, translation t_p, and scale s_p, along with the query position x. It outputs signed distances ŝ_p for all parts at the queried position, which may be combined into the global signed distance ŝ. Our decoder alternates between updating each part independently and sharing information across parts, allowing them to adapt to one another while preserving modularity. This is implemented through a sequence of lightweight convolutions applied along rows and columns of the part feature matrix.

Results

Reconstruction

Reconstruction of test shapes. For part-based methods, we color each part with a different color and translate the helix outside of the mixers for visualization.

Generation

Shape generation and pose-conditioned generation (Ours^†) where part latents are generated based on the poses' coarse description of the shape (left image for each pair). When possible, the helix is translated outside of the mixer for visualization.

Manipulation

PartSDF represents 3D shapes as assemblies of controllable parts. Each component can be moved, reshaped, or replaced independently, while the model preserves overall consistency. This enables intuitive part-based editing for design and optimization tasks.

Optimization

Using PartSDF as a differentiable shape representation, complex designs can be optimized directly in 3D. Here, the car body is automatically refined to minimize drag, while the wheels remain fixed, showcasing part-aware optimization with overall consistency.

We show individual parts (left) and the surface pressure (right).

BibTeX

If you find our work useful, please cite it:

@misc{talabot2025partsdf,
    title={PartSDF: Part-Based Implicit Neural Representation for Composite 3D Shape Parametrization and Optimization}, 
    author={Nicolas Talabot and Olivier Clerc and Arda Cinar Demirtas and Hieu Le and Doruk Oner and Pascal Fua},
    year={2025},
    eprint={2502.12985},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2502.12985}
}