Diffusion Models
Restoring a Ruined Painting
7 min read
Restoring a Ruined Painting
Imagine taking a crystal-clear painting, slowly splashing paint on it until it's pure noise — and then learning to reverse that process.
An art restorer studies thousands of ruined paintings being restored back to clarity. They learn the patterns of how noise becomes structure. Diffusion Models train by adding random noise to images step by step until nothing recognisable remains, then learning to reverse that process. When you give it a text prompt, it starts from pure noise and gradually sculpts an image by reversing the noise.
In Plain English
Diffusion Models generate images by learning to remove noise. They're trained by taking clear images, progressively adding random noise until unrecognisable, and then learning to reverse that — going from noise back to a coherent image guided by your text prompt.
The Technical Picture
Diffusion Models are probabilistic generative models that learn a data distribution by modelling a Markov chain of diffusion steps. The forward process adds Gaussian noise; the reverse process uses a neural network (typically a U-Net) to denoise, conditioned on a text embedding from a CLIP or similar encoder.
Real-World Examples
- Midjourney and DALL-E 3 use diffusion models to generate images
- Adobe Firefly's 'Generate Image' feature
- Stable Diffusion (open-source) running locally on your laptop
Diffusion Models sculpt images by starting from pure noise and gradually removing it — guided by your words.