Over the last few months, open-source image and video models have gotten really good. The part that still feels unsolved (at least for me) is creating long-form generative content with real control, without everything collapsing into a single giant prompt or a VRAM-melting pipeline.

ComfyUI workflows are already a great abstraction, but once you start doing things like first/last frame control, LoRAs, WAN, or multi-stage generation, you quickly end up juggling lots of separate workflows and manual hand-offs. That’s manageable for a couple runs, but it doesn’t scale if you’re trying to build longer sequences or iterate systematically.

So I built a small orchestration layer that treats ComfyUI workflows as the atomic unit. The idea is:

workflows stay simple and composable

you can register multiple ComfyUI backends (local or remote)

the orchestrator handles sequencing and parallel execution

everything runs locally/offline, no hosted services

In practice, this makes it much easier to generate long-form content by chaining controlled steps (e.g. generate → refine → constrain → extend) instead of pushing everything through one monolithic graph.