diffusion

Definition

A generative approach that learns to reverse a gradual noising process: start from noise and denoise step by step until a sample emerges. Splitting generation into many small, individually easy steps makes training unusually stable, which is why diffusion dominates image, video, and audio synthesis. The cost is iterative sampling — many forward passes per sample — motivating extensive work on step reduction and distilled samplers.

Explainers using this term

• Paper Walkthrough: On-Policy Self-Distillation in Diffusion Models — Turning Reward into a Target You Can Aim AtOn-Policy Self-Distillation in Diffusion Models
• Build Your Own Diffusion Model — Starting from MNISTDenoising Diffusion Probabilistic Models
• CFG and Samplers — What the "Strength" Knob in Generative AI Really DoesClassifier-Free Diffusion Guidance