Jargon:Diffusion model
1 min read
Say this at the party
“A diffusion model makes pictures by starting with pure static and cleaning it up, step by step, until an image appears.”
In plain English
To train one, you take real pictures and slowly add random noise until they look like TV static. The AI learns to undo each little step.
To make something new, it starts from fresh static and removes noise bit by bit, guided by your words. A picture slowly comes into focus.
This is the idea behind many image tools you have heard of, such as Stable Diffusion. It is used for video and sound too.
Where it came from
Diffusion models were introduced in 2015 by Jascha Sohl-Dickstein and colleagues, borrowing ideas from physics about how things spread out. They took off after a 2020 paper by Jonathan Ho and colleagues made them work well, and a 2021 paper led to Stable Diffusion.
See it in action
Ade types "a fox in a raincoat reading a newspaper" into an image app. Within seconds, a blurry mess of static sharpens into exactly that.
What people get wrong
- It doesn't cut and paste bits of existing pictures. It builds a new image from noise, using patterns it learned from training images.
- Not every AI image tool uses diffusion. Some use different methods, and many mix several together.
Go deeper
- PaperSohl-Dickstein et al., 2015Deep Unsupervised Learning using Nonequilibrium ThermodynamicsThe original diffusion model paper, inspired by physics.
- PaperHo, Jain and Abbeel, 2020Denoising Diffusion Probabilistic ModelsThe paper that made diffusion images good enough to take off.
- PaperRombach et al., 2021High-Resolution Image Synthesis with Latent Diffusion ModelsThe research behind Stable Diffusion.
- ArticleWikipediaDiffusion model - WikipediaA thorough overview, if you're happy to skip the maths.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.