Comment by armcat

If you know how an LLM could have an output image head, that is purely generative without using diffusion, and not a classification/selection head, I’m really interested in learning more. I think you are onto something very new and novel here