Unpaired 3D Shape-to-Shape Translation via Gradient-Guided Triplane Diffusion.

Area of Science:

• Computer Vision
• 3D Shape Analysis
• Generative Models

Background:

• Unpaired shape-to-shape translation transforms 3D shapes without paired data.
• Existing GAN-based methods struggle with diversity and quality due to "mode collapse".

Purpose of the Study:

• To develop a new method for high-fidelity, diverse unpaired shape-to-shape translation.
• To address the limitations of existing generative adversarial network (GAN) approaches.

Main Methods:

• Utilizes a triplane diffusion model to factorize 3D objects into triplane representations.
• Applies a forward diffusion process to smooth shape structures and a reverse diffusion process for denoising.
• Incorporates gradient-based guidance mechanisms to enhance translation faithfulness.

Main Results:

• The triplane diffusion model achieves superior shape fidelity and high-quality results.
• Demonstrates improved generation diversity compared to state-of-the-art baselines.
• Successfully translates shapes across different domains while preserving structural similarities.

Conclusions:

• The proposed triplane diffusion model offers a robust solution for unpaired shape-to-shape translation.
• Gradient-based guidance effectively improves the accuracy and quality of shape transformations.
• This method advances the field of 3D generative modeling for shape manipulation.