Bidirectional Beta-Tuned Diffusion Model.

Area of Science:

• Machine Learning
• Generative Modeling
• Deep Learning

Background:

• Diffusion models generate high-quality samples but uniform training is suboptimal.
• Analysis reveals non-uniform distribution variations in the forward diffusion process, with rapid initial changes.

Purpose of the Study:

• To theoretically analyze the forward process of diffusion models.
• To propose a novel diffusion model training strategy that addresses suboptimal uniform timestep sampling.

Main Methods:

• Comprehensive theoretical analysis of the forward diffusion process.
• Introduction of the Bidirectional Beta-Tuned Diffusion Model (BB-TDM).
• Leveraging the Beta distribution for timestep sampling in BB-TDM.

Main Results:

• Initial distributions rapidly converge to Gaussian, diminishing differences early in the process.
• Uniform timestep sampling fails to capture these dynamics effectively.
• BB-TDM enhances separation between initial distributions and aligns sampling with forward process properties.

Conclusions:

• The BB-TDM effectively moderates convergence speed and improves diffusion model training.
• Experiments confirm BB-TDM's efficacy across benchmark datasets and diffusion models.