What is a diffusion model?

A diffusion model is a generative model that learns to create data by reversing a gradual noising process.

During training, noise is progressively added to real data until the original structure is mostly destroyed. The model then learns how to reverse that process step by step.

Core idea

A simplified view is:

Real data → Add noise gradually → Nearly pure noise

Then generation works in the opposite direction:

Random noise → Remove noise step by step → New sample

How does the model learn?

The model is trained to predict and remove the noise added at different stages.

By learning this denoising process, it gradually learns the structure of the training-data distribution.

What can diffusion models generate?

Diffusion models are commonly used for:

• Image generation
• Image editing
• Inpainting
• Super-resolution
• Audio generation
• Some video and multimodal generation tasks

Why are diffusion models popular?

They can generate very high-quality and diverse outputs, especially for images.

Compared with GANs, diffusion models are often more stable to train, although generation can require many iterative denoising steps and therefore be computationally expensive.

Main limitation

Sampling is usually slower than a single forward pass because the model repeatedly refines noise over multiple steps.

For the broader category, see “What is a generative model in deep learning?”. The full generation process is covered separately in “How do diffusion models work?”.

In short, a diffusion model learns how to reverse a noise-adding process so it can transform random noise into realistic new data.