Wiki · AI & Machine Learning · concept
Diffusion model
A generative model that learns to turn pure noise into a sample by reversing a gradual noising process, one denoising step at a time.
A diffusion model generates by learning to undo noise. You take real data and slowly corrupt it with random noise until nothing is left but static. Then you train a model to run that process backwards — to look at a noisy thing and guess what the slightly-cleaner version underneath looked like. Once it can do that, you hand it pure noise and let it denoise, step by step, into a brand-new sample that never existed but looks like it belongs.
How it works
There are two halves. The forward process is fixed and dumb: start with a real image, protein, or molecule and add a little random noise, then a little more, then more, over many steps until it dissolves into featureless static. Nothing is learned here — it’s just controlled destruction. The reverse process is where the model lives. It’s trained on every step of that decay to answer one question: given this noisy version, what noise do I subtract to get closer to something real? Think of a sculptor who has watched thousands of statues slowly buried under sand, and has learned to brush the sand away in exactly the right order. At generation time you start from a fresh pile of sand — random noise, no statue underneath — and the model brushes, guided at each step by what it learned, until a coherent shape emerges. Steer that denoising with a text prompt or a target and you get controllable generation.
Why it matters (for bio × AI)
Diffusion is the generative engine under modern structural biology. It’s how AlphaFold3, Boltz, and Chai place atoms when they co-fold proteins with ligands, DNA, and drugs; how RFdiffusion hallucinates entirely new protein backbones and custom binders that grip a chosen target; and how generative-chemistry models sketch novel small molecules. The same trick that paints an image — start from noise, denoise toward something plausible — turns out to design a protein.
Related
See transformer for the architecture that usually powers the denoiser, and latent space for the learned map many diffusion models denoise within rather than in raw coordinates.