Diffusion Model
A class of generative AI model that learns to create new data (images, proteins, molecules, weather states) by learning to reverse a gradual noising process — the basis of tools like Stable Diffusion, RFdiffusion, and GenCast.
What it means
A diffusion model is a type of generative model trained to produce new data by learning to reverse a noising process. During training, the model is shown how clean data (a protein structure, a weather state, a molecule) gradually degrades into random noise, step by step. It then learns to run that process in reverse: starting from pure noise, it iteratively denoises toward a plausible output.
The result is a model that can generate new examples of the kind of data it was trained on — not by memorizing and replaying training examples, but by learning the underlying statistical structure of the data distribution.
Why it matters for research
Diffusion models have become the dominant approach in scientific generative AI across several domains:
- Protein design: RFdiffusion uses diffusion to generate protein backbone structures that can be designed to fold stably or bind a target molecule
- Weather forecasting: GenCast uses a diffusion-based approach to generate ensemble forecasts — many plausible future weather states — rather than a single deterministic prediction
- Drug discovery: Several molecular generation tools use diffusion models to propose candidate molecules with specified properties
- Structure prediction: DiffDock uses diffusion to model how a drug molecule docks into a protein binding site
The key advantage over earlier generative models (like GANs) is stability and diversity of outputs. Diffusion models are less prone to mode collapse (generating only a narrow range of outputs) and tend to produce higher-quality results across a wider distribution.
The iterative denoising process
At inference time, a diffusion model:
- Starts from a sample of pure random noise
- Applies a learned denoising step, moving the sample slightly toward more structured data
- Repeats this many times (typically tens to hundreds of steps)
- Produces a final output that looks like a plausible example from the training distribution
Because each run starts from different random noise, running the model multiple times produces different outputs — this is how ensemble models like GenCast generate a distribution of forecasts rather than a single prediction.
Related terms
- Latent Space — the compressed representation space many generative models work in
- De Novo Design — the protein/molecule design application of generative models
- Simulation-Based Inference — a related approach for scientific inference under uncertainty