Can someone please explain how I can run this on my computer but something like GPT3 is too computational intensive to do the same? Isn’t text easier than images?
Stable Difussion isn't so heavy... mostly you are limited by how many steps you want to do.
An image without “crisp” pixels that create lines is acceptable. A basketball with an extra/missing blue pixel is acceptable.
With words, they either haven’t invented Diffusion yet, or the nature of the problem is too hard for Diffusion.
With text you can’t have the model Diffuse to “teli ne abouf a fdog?” “The dod is graen and has stix leg5”. It’s just too obviously wrong.
Meanwhile, the “fuzziness” allows Diffusion models to be smaller, compared with models that need precision.