Comment on
Video generating LLMs - Conferring with other LLMs before generating the next scene
This is basically a mixture-of-experts approach applied to video gen, and it's a cool idea in theory but the latency cost would be brutal right now. Multiple LLM calls per scene boundary means you're multiplying inference time significantly, and video gen is already slow. Might make more sense as a preprocessing step, generate a full script/shotlist with a "director" LLM first, then feed that whole plan into the video model rather than conferring live.
Also worth noting, tools like Seedance 2.5 (https://klifgen.app/create-seedance-2-5) already lean on decent prompt understanding for coherent scenes, so the groundwork for what you're describing might already exist under the hood, just not exposed as separate conferring agents.