Denoising Diffusion Probabilistic Models

Abstract

We present high quality image synthesis results using diffusion probabilistic models, a class of latent variable models inspired by considerations from nonequilibrium thermodynamics. Our models produce samples that are competitive with state-of-the-art GANs while enjoying desirable properties such as distribution coverage and a stationary training objective.

Key Findings

• 1Demonstrated that diffusion models can generate high-quality images competitive with GANs
• 2Introduced a simplified training objective based on denoising score matching
• 3Showed better mode coverage and training stability compared to GANs
• 4Established the theoretical framework for modern diffusion-based generation
• 5Achieved state-of-the-art FID scores on image generation benchmarks

Impact & Significance