[RestlessAPI]

I wonder if this works for Stable Diffusion as well...

[acheong08]

From what I understand, Stable Diffusion starts out with random noise that is then step by step made into the final image. Randomness is there at the beginning.

Current LLMs first find the n most likely next token and only then gets randomness on the choice among the top-n. This injects randomness to the initial search for the top-n.

So I think, from my ignorance, that Stable Diffusion is already doing this in some sense

[samus]

There are transformer-based approaches to vision models. It should work for those.