Neural Wavelet-domain Diffusion for 3D Shape Generation, Inversion, and Manipulation

TL;DR

This paper introduces a wavelet-domain diffusion method for 3D shape generation, inversion, and manipulation using neural networks that operate on a continuous implicit wavelet representation, enabling diverse and detailed shape control.

Contribution

It proposes a novel wavelet-based implicit representation and a diffusion model for 3D shape synthesis, inversion, and manipulation, advancing the state-of-the-art in shape modeling.

Findings

Outperforms existing methods in shape generation quality.

Enables detailed shape manipulation and inversion.

Demonstrates strong quantitative and qualitative results.

Abstract

This paper presents a new approach for 3D shape generation, inversion, and manipulation, through a direct generative modeling on a continuous implicit representation in wavelet domain. Specifically, we propose a compact wavelet representation with a pair of coarse and detail coefficient volumes to implicitly represent 3D shapes via truncated signed distance functions and multi-scale biorthogonal wavelets. Then, we design a pair of neural networks: a diffusion-based generator to produce diverse shapes in the form of the coarse coefficient volumes and a detail predictor to produce compatible detail coefficient volumes for introducing fine structures and details. Further, we may jointly train an encoder network to learn a latent space for inverting shapes, allowing us to enable a rich variety of whole-shape and region-aware shape manipulations. Both quantitative and qualitative experimental results manifest the compelling shape generation, inversion, and manipulation capabilities of our approach over the state-of-the-art methods.