Efficient and Scalable Point Cloud Generation With Sparse Point-Voxel Diffusion Models

Area of Science:

• Computer Vision
• Machine Learning
• 3-D Modeling

Background:

• Generative modeling for 3-D shapes is crucial for various applications.
• Existing methods often struggle with balancing generation quality, diversity, and speed.
• Diffusion models show promise but can be computationally intensive.

Purpose of the Study:

• To propose a novel point cloud U-Net diffusion architecture for efficient and high-quality 3-D generative modeling.
• To evaluate the proposed architecture's performance on unconditional and conditional shape generation, completion, and super-resolution tasks.
• To demonstrate the model's speed and scalability.

Main Methods:

• A dual-branch architecture combining point and sparse voxel representations.
• Utilizing a U-Net diffusion framework for generative tasks.
• Extensive evaluations on benchmarks like ShapeNet for various 3-D generation and processing tasks.

Main Results:

• The fastest variant surpasses non-diffusion methods in unconditional shape generation.
• The largest model achieves state-of-the-art results among diffusion methods, with a 70% runtime of prior state-of-the-art.
• The model shows scalability to larger datasets and excels in conditional generation, completion, and super-resolution.

Conclusions:

• The proposed point cloud U-Net diffusion architecture is a state-of-the-art solution for 3-D generative modeling.
• The architecture offers a compelling balance of generation quality, diversity, and speed.
• The model's versatility across multiple 3-D tasks highlights its potential impact.