[johntb86]

Meta has experimented with a hybrid mode, where the LLM uses autoregressive mode for text, but within a set of delimiters will switch to diffusion mode to generate images. In principle it's the best of both worlds.