Definition
A diffusion model is a generative model trained to reverse a gradual noising process: during training the model learns to predict and remove noise at each step; at generation time it starts from pure noise and denoises step by step to produce images, audio and other samples.
How it works
The process has two directions: a fixed forward process that gradually turns data into noise, and a learned reverse process that restores the data step by step. Diffusion models are known for high output quality and stable training, at the cost of multi-step inference that is usually slower than single-pass generators.
Usage
Diffusion models are widely used for text-to-image generation, video and audio synthesis, image restoration and super-resolution, and they underpin many modern speech synthesis systems.
Related and confusable terms
LoRA is often used as an efficient fine-tuning method for diffusion models: one is the model itself, the other is an adaptation technique, and they frequently appear together in community model sharing.