A diffusion model is a generative model that learns to create data by reversing a gradual noising process.
A diffusion model learns how to turn noisy data into useful, structured data.
During training, noise is gradually added to examples. The model learns how to predict or remove that noise, so generation can start with noise and repeatedly make it more organized.
Starting with a television screen full of static and gradually turning the noise into a recognizable image.
This is only a learning analogy: a diffusion model is not literally cleaning a television screen.
Creates an image from a text prompt or other conditioning signal.
Uses controlled noise and denoising to change part of an existing image.
Produces sequences of frames while aiming for temporal consistency.
Diffusion: A generative modeling family based on noising and denoising.
Transformer: A neural-network architecture based on attention.
Generative AI: A broader capability or category for creating new content.
Diffusion: One family that can support generative AI.
Diffusion: Can be used with images, but also with other data types.
Image model: A broader description based on data or task.
Diffusion is one family within generative modeling. GAN and VAE pages are not linked because they are not currently available.
Explore next
A diffusion model learns to turn noise into structured data through repeated denoising.
A verified video explainer will be added when a real asset is available.