PartMat: Material-Aware 3D Part Decomposition with a Single Global Latent

PartMat is a new material-aware 3D part decomposition pipeline built around a single global latent, rather than generating each part independently. That matters because a lot of current part-based 3D systems still optimize for semantic structure—chair legs, cabinet doors, and so on—while production work often needs editable splits along real material boundaries like wood, fabric, and metal.

The core idea is PartVAE, which learns a unified representation of the whole object and decodes all material parts in one forward pass. In practical terms, that means inference cost no longer grows linearly with part count, which is a meaningful efficiency win for tools that need to handle complex assets with many subcomponents.

On top of that representation, the pipeline trains a diffusion model for part generation and then refines it with reinforcement learning to improve material assignment and reduce overlaps. A sparse-voxel flow-matching stage with part attention is then used to recover finer geometric detail after decomposition.

The reported result is a system that outperforms existing baselines on material-aware decomposition accuracy while keeping geometric quality competitive. For game teams, the interesting angle is not just cleaner asset segmentation, but a path toward faster, more editable 3D content pipelines for interior props, modular environment assets, and any workflow where material edits matter as much as shape.