Checked 25 September 2026 · vLLM-Omni client scope
The H3 server can take first and last frames. vLLM-Omni’s current ComfyUI client cannot send both.
The update: vLLM-Omni’s H3 recipe documents ordered first-and-last-frame conditioning. Its current API/ComfyUI RFC says the server supports several H3 input modes, while the current ComfyUI client accepts one FL2VA frame and limits some reference combinations. Do not buy a transition workflow from an endpoint label.
Server support and client support are different facts
Project says: the serving recipe shows an FL2VA request with ordered first and last images. The current RFC says the server supports T2VA, first-frame FL2VA, mixed Ref2VA, repeated references, reference ordering, and H3 reference limits.
Project says: the current vLLM-Omni ComfyUI client supports only one FL2VA frame, caps each reference modality at two inputs, and rejects some valid mixed-reference combinations. The RFC lists client work for separate first and last frames and broader reference support. It is an open plan, not proof that those workflows are ready.
What this changes for a continuity build
A first/last-frame transition needs both keyframes to reach the server. A server that accepts them does not make a client that submits one frame into a two-keyframe workflow. Keep these labels separate: server capability, client capability, and a recorded output.
This also limits the two-RTX-5090 result. The documented consumer-GPU record is one full-H3 T2VA job, not a first/last-frame result. It cannot validate a ComfyUI transition path.
FastH3 is a server choice, not a UI toggle
Project says: FastH3 uses one selected adapter fused when the server starts. The current Remote LoRA control does not switch it per request, and the RFC’s planned FastH3 node is limited to the adapter’s four-step T2VA contract.
Choose the task and server before you choose the Comfy graph. Do not treat a local LoRA-looking control as proof that a remote FastH3 service changed variants or accepted a frame-anchored task.
Smallest honest check
If you are authorised to test the route, pin the server and client revisions. Record one text job, one first-frame job, one first-and-last-frame job, and one richer reference combination. Save the outgoing request, returned MP4, audio track, and client error. A control that is missing, a request that is rejected, and a complete video are different results.