Denoising Diffusion Probabilistic Models

TL;DR

Frames image generation as learning to reverse a gradual noising process: start from pure noise and denoise step by step into a sample.

Why it matters

This paper turned diffusion into a practical, high-quality generative method and set off the wave of modern image and video generators. It reframed generation as iterative denoising.

Key ideas

• A forward process slowly adds Gaussian noise to data over many steps.
• A network learns to predict and remove that noise, reversing the process.
• Sampling starts from noise and denoises into a realistic image.
• Stable to train and produces high-fidelity samples.

Related concepts