Disentangle-and-Diffuse: Structure-Aware Modeling with Diffusion-Based Generation

We introduce Disentangle-and-Diffuse, a framework for structure-aware 3D shape generation that combines automatic part decomposition, invariant/equivariant feature encoding, and diffusion-based part synthesis. The pipeline first segments a 3D shape into semantic parts, then encodes each part with geometry features that are invariant to global rotations and pose features that transform equivariantly with the object orientation. A dual-stream transformer fuses local part context and global shape context to condition a part diffusion model, and the generated parts are assembled with predicted pose features. Experiments on challenging benchmarks show that our method improves structural consistency and geometric quality over the evaluated baselines, while enabling flexible shape manipulation and interpolation. These results support the value of combining structural and feature-level disentanglement for controllable 3D generative modeling.