Protoface MiniMax H3 FastH3 (Preview) · Generate Video

The speed tier, and a note on the name

This is MiniMax H3 in its faster configuration - note the category (Protoface/Video/protoface rather than /minimax) and the (Preview) in the display name. Protoface lists it on their own shelf, and preview is honest labelling: it's the tier you reach for when you need a take now and don't need the flagship's full reference control.

What you give up versus the general H3 node: the reference sockets. There's no reference, reference_video, audio, first_frame or last_frame here - no image conditioning at all. It's text to video, full stop. What you keep: native synchronised audio, duration_seconds from 5 to 15, resolution tiers, aspect ratio, and an actual seed, which the general H3 node doesn't expose.

So: text-only, faster, reproducible. That's a coherent little node.

The inputs

prompt is the shot. With no reference media, wording is everything - subject, motion, camera, and the sound you want, because the audio is generated from the same prompt.

quality defaults to 480p and its tooltip makes a point the whole pack keeps making: "each tier is an exact canvas, not a label." That's a hint that an upscaled tier isn't a different render, and it's your main cost dial. Start at 480p while you're working out the shot.

duration_seconds runs 5 to 15 with a default of 8. Longer means more money; it also means more chances for the motion to drift, so short-and-good usually wins.

aspect_ratio defaults to 16:9 as a picker - no free-text values here.

generate_audio defaults to true, and its tooltip is the most useful sentence on the node: FastH3 emits synchronized audio from the same pass, and disabling it does not reduce cost. So there is no cheap silent mode. If you don't want the audio, mute it downstream and stop thinking about it as a savings lever.

enhance_prompt is off by default. Per its tooltip it rewrites your prompt into the engine's native grammar, best effort, and passes the prompt through unchanged if the rewrite is unavailable or would change the direction you asked for. Nothing to fear, occasionally helpful.

seed is a dynamic combo here - -1 for random, a number for a repeatable attempt. Handy when you have a good frame and want to nudge one word.

generation_id and request_scope: the same booking pair as every node. Fixed ID resumes the same paid job; increment for a new generation. Scope is auto in the UI, required by headless clients.

Outputs

video (VIDEO, list), files (PROTOFACE_FILE, list - the originals, and with native audio in play you want the original container, so Protoface Save File rather than a re-encode), run id for resume and support.

Install

cd ComfyUI/custom_nodes
git clone --branch v0.1.0 --depth 1 https://github.com/protoface-ai/protoface-comfyui.git protoface
python -m pip install -r protoface/requirements.txt

ComfyUI 0.41.0+, Python 3.10+, deps httpx, jsonschema, Pillow, av>=14. Or Manager → search Protoface → protoface-nodes → restart → reload. Protoface button, lower-right, then your key. Memory-only: re-enter after a server restart, or use PROTOFACE_API_KEY_FILE.

Where people get bitten

Expecting image-to-video. There's no image socket on this node, and no amount of hunting will produce one - use the general H3 node or the Hailuo nodes for that.

Expecting a silent cheap mode. generate_audio off costs the same.

And if a job is slow, wait or resume rather than re-queue: preview-tier models share a queue, timeouts leave the job recoverable, and the fix is the same generation_id plus a requeue, or Protoface Resume Run with the run id. Cancelling mid-flight is best effort and carries no refund. Also leave ~/.cache/protoface-comfyui alone - it holds the journal that makes recovery cheap.