The text-to-image model is an important component, but the current model in use is IMO good enough. My view for this project is that the internal monologue is more important than the output, so my wish is instead for a better open-weight LLM.
LLM: Mixtral-8x7B
text-to-image: one of the leading commercial models, whose TOS I may or may not be violating.
Mixtral is great. I assume you saw the DBRX and new larger Mixtral release that just came out over the last few days.
I did! I want to switch to Mixtral-8x22B, time permitting. During the development of Stream of Consciousness I already swapped LLMs twice. This space is moving incredibly fast.