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

TL;DR

PartSDF tackles the lack of part-aware structure in existing 3D INRs by introducing per-part latent codes $\mathbf{z}_p$ and poses $\mathbf{p}_p$ decoded via a shared cross-part INR. The global SDF is obtained as $\hat{s}=\min_p \hat{s}_p$, enabling seamless composition of independent parts while maintaining coherence. Training relies on region-based supervision from the global SDF and a non-intersection loss to handle non-watertight parts, enabling learning from non-watertight segmentations. The results show improved reconstruction, diverse part-aware generation, and effective part manipulation and optimization, highlighting PartSDF as a practical shape prior for engineering design.

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. Code available at https://github.com/cvlab-epfl/PartSDF.

Figures (22)

• Figure 1: Part-based implicit representation. PartSDF is a simple and modular approach to representing composite 3D shapes for many different purposes: (a) Shape generation, possibly conditioned on a part layout. (b) Part-based manipulation. (c) Part-aware optimization. In this example, the car body, shown in red, is optimized to reduce aerodynamic drag while maintaining the shape and position of the wheels fixed.
• Figure 2: PartSDF pipeline. (a) Our model's core is a part auto-decoder $f_\theta$ that takes as input part latents $\mathbf{z}_p$ and poses expressed in terms of a quaternion $\mathbf{q}_p$, translation $\mathbf{t}_p$, and scale $\mathbf{s}_p$, along with the query position $\mathbf{x}\in\mathbb{R}^3$. It outputs signed distances $\hat{s}_p$ for all parts at the queried position, which may be combined into the global signed distance. (b) A secondary model may be used based on the task at hand, such as encoders to map a given modality, e.g., point clouds, to part latents and poses, or a diffusion model to generate them from noise.
• Figure 3: Cross-part adaptation in PartSDF. 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.
• Figure 4: Supervision for non-watertight parts. (a) Semantic parts of a chair, (b) a 2D slice of its signed distances (red/blue) with parts highlighted in color, and (c) the specific regions of space where each part is supervised. This enables training without requiring parts to be watertight.
• Figure 5: Shape reconstruction. Reconstruction of test shapes with all models. For part-based methods, we color each part with a different color and translate the helix outside of the mixers for visualization.