A plain summary, so you can get the gist here without leaving.
In 2020, three researchers showed a clean way to generate images by teaching a model to reverse noise: start from pure static and remove a little noise at a time until a real-looking picture appears.
What it is
This paper, by Jonathan Ho, Ajay Jain, and Pieter Abbeel, develops diffusion models for generating images. The training has two directions. Going forward, you take a real image and gradually add random noise until it becomes meaningless static. Going backward, the model learns to undo that, removing noise step by step.
Once trained, the model can start from fresh random noise and walk the reverse path, cleaning it up bit by bit until a brand new, coherent image emerges.
The core idea
It is easier to learn to remove a small amount of noise than to paint a whole image in one go. Diffusion breaks the hard problem of creating an image into many small, gentle steps, each just a little cleanup.
Because the model only ever learns to denoise slightly, the task at each step is manageable, and the steps add up to something remarkable. Many tiny corrections turn random static into a detailed picture.
Why it matters
Diffusion became the engine behind much of today's AI image generation. The step-by-step denoising approach produces high quality and varied results, and it has spread to audio, video, and beyond.
If you have used a tool that turns a text description into a picture, there is a good chance diffusion is doing the work underneath. For builders, this paper is the accessible foundation of modern generative imaging.
- Published in 2020 by Ho, Jain, and Abbeel.
- Trains by adding noise to images, then learns to reverse it.
- Generates new images by denoising random static step by step.
- Breaks a hard creation task into many small, manageable steps.
- Became a core method behind modern AI image generation.
Ho, Jain & Abbeel
New to this? Come build with us.
Reading is good. Building with people is better. Our drop-ins are free and open to total beginners.