> so they can't be integrated That's overly pessimistic. Diffusion models take an input and produce an output. It's perfectly possible to auto-regressively ana

> so they can't be integrated That's overly pessimistic. Diffusion models take an input and produce an output. It's perfectly possible to auto-regressively analyze everything up to the image, use that context to produce a diffusion image, and incorporate the image into subsequent auto-regressive shenanigans. You'll preserve all the conditional probability factorizations the LLM needs while dropping a diffusion model in the middle.