vLLM-Omni ComfyUI integration for MiniMax-H3 (T2VA / FL2VA / Ref2VA)
Sharing an early ComfyUI integration for serving MiniMax-H3 through vLLM-Omni:
PR: https://github.com/vllm-project/vllm-omni/pull/5756
The PR extends vLLM-Omni's existing ComfyUI-vLLM-Omni custom nodes with MiniMax-H3 video-and-audio generation. The currently wired workflows are:
- T2VA: prompt only
- FL2VA: prompt plus a first-frame image
- Ref2VA: one image plus one audio reference
- Ref2VA: one or two reference videos
The ComfyUI client automatically selects t2va, fl2va, or ref2va from the connected inputs and exposes H3's flow_shift and audio_flow_shift. It talks to the OpenAI-compatible vLLM-Omni /v1/videos service, so ComfyUI can run on a different machine and in a separate environment from the H3 GPU server.
This complements the recently merged native ComfyUI MiniMax-H3 implementation rather than replacing it. Native ComfyUI currently exposes richer reference combinations; the vLLM-Omni path is aimed at users who want:
- a dedicated local or remote inference server;
- isolated ComfyUI and model-serving environments;
- multi-GPU serving, VAE patch parallelism, or CPU-offload deployment on the server side;
- one API backend shared by ComfyUI and programmatic clients.
To try the open PR, copy apps/ComfyUI-vLLM-Omni into ComfyUI/custom_nodes/ComfyUI-vLLM-Omni, start the appropriate H3 FL2VA or Ref2VA checkpoint with vllm serve ... --omni, then use the included vLLM-Omni Video Generation template. The model card's vLLM recipe remains the best source for the current H3 server command.
Feedback from the MiniMax team and H3 users would be especially helpful on reference-input semantics, recommended quality validation cases, and which additional Ref2VA combinations should be prioritized next.