David Forsyth

 

 

David A. Forsyth

 

 

 

 

Talk: What diffusion models do:  not what it says on the box, and worth fixing

Diffusion models generate images from noise, and are now the backbone of image generation activities. I show that: generative models "know" properties of scenes that aren't in their training data; don't "know" other properties of the natural world; and leave distinctive fingerprints in the images they generate. This doesn't matter if the application of the models is to make commercial art cheaper or public discourse nastier, but matters a lot if the models are used to solve, say, inverse problems.

Workshop Home Page

Return to the TAU-UIUC Workshop Home Page

Home→