Part-Aware Shape Generation With Latent 3D Diffusion of Neural Voxel Fields

Area of Science:

• Computer Vision
• 3D Shape Generation
• Deep Learning

Background:

• Generating high-quality 3D shapes with precise structural understanding remains a challenge.
• Existing methods often struggle with part-aware decomposition and detailed texture synthesis.

Purpose of the Study:

• To introduce a novel latent 3D diffusion model for generating neural voxel fields with part-aware structures and high-quality textures.
• To improve the accuracy and quality of 3D shape generation by incorporating part-aware information throughout the diffusion process.

Main Methods:

• A latent 3D diffusion process for neural voxel fields is introduced, integrating part-aware information.
• A part-aware shape decoder is utilized to guide the integration of part codes into neural voxel fields.
• Part-aware learning establishes structural relationships for texture generation in similar regions.

Main Results:

• The proposed method demonstrates superior generative capabilities in part-aware shape generation across eight data classes.
• It outperforms existing state-of-the-art methods in generating accurate part decomposition and high-quality rendering.
• Successful image- and text-guided shape generation showcases multi-modal potential.

Conclusions:

• The novel latent 3D diffusion model effectively generates neural voxel fields with precise part-aware structures and high-quality textures.
• The approach significantly advances the state-of-the-art in part-aware 3D shape generation and rendering.
• The model shows promise for multi-modal guided 3D content creation.