OpenH3-IR Main

The whole pack, in one node. You type one plain sentence into OpenH3-IR Main, hit queue, and out the other side come the model, conditioning, latent and both VAEs for a finished MiniMax H3 job. No prompt copy-pasting from a chat window, no <Picture 1> bookkeeping, no separate service to keep alive. This is the node you'll actually live in.

Here's why it exists. MiniMax H3 - the open 33B omni-modal video model, ~42.5 GB, generates 4–15 second clips with native stereo audio - doesn't take a prompt the way Stable Diffusion does. It consumes a structured Context-IR document: a formalized brief with shots, references, dialogue and sound all spelled out. Writing that document by hand is miserable, and getting an LLM to write it in a chat window means copying text back into ComfyUI and hoping the model names your references right. OpenH3-IR is an open implementation of that Context-IR layer, and Main is its ComfyUI front door.

How it works

Main runs the OpenH3-IR compiler, which asks your language model (configured on OpenH3-IR Setup) to write the brief from your sentence, validates the result, picks the right H3 checkpoint for the job, loads the encoder and both VAEs, and hands everything to the render side. This is the "LLM as prompt builder" pattern, but with teeth: the output has to be a valid Context-IR document, and anything that can't survive validation is refused and rewritten rather than silently accepted.

The prompt syntax is deliberately tiny. Plain prose for what happens, @ to name something in the media tray, and @speaks("exact words") to lock dialogue. That locked span has to come back word for word and mark for mark - a brief that rewords it is thrown out and written again.

The inputs that matter

• intent - the one sentence. Not a tag list and not a shot breakdown; the compiler writes those. Say the action and the beat you care about.
• seconds - the only place length is set, used by both the brief and the latent. H3 renders on a 17k+5 frame grid, so this snaps up: ask for 10 and you get 10.125. 8.0 is the only whole second on the grid. The trained band is 5.167–15.083 s; outside it renders anyway, untested and slower, and the report tells you so.
• creativity - how much the writer may add where your sentence is silent: restrained, balanced, bold, extreme. It controls three things only - music, a spoken line, and text in the frame. It never decides shot count.
• no music - H3 writes sound in the same pass as the picture, so silence is a decision. This turns off music only; ambient and physical sound still get written.
• shots - auto lets the writer cut, or pin 1–10. Every shot needs 1.2 s, so a count that can't fit the duration is refused with the arithmetic rather than silently changed.

The optional inputs are worth a glance once you're past your first render: megapixels (0 = H3's native 768 short edge; more is sharper, slower, and eats VRAM), seed (the brief seed - changes the writing, not the sampler), and sizing (match fits each picture to the render, max keeps the picture's own size for stronger identity and costs more).

The outputs

Main gives you model (already the right checkpoint - feed your LoRAs and sigma shift from here), positive conditioning, latent (empty, already the length the brief was written for), vae and audio_vae (no loader boxes needed), plus two read-outs: prompt, the compiled brief, and report, a plain-words account of the real length, what each @ became, and every file loaded. Wire report into ComfyUI's Preview as Text node; you'll want it when something's off.

Your sampler, LoRAs, steps, decode and save all stay yours. The pack deliberately starts close to ordinary H3 settings, including the beta scheduler - Comfy-Org's reference-to-video guidance says beta or normal beats simple on reference-heavy prompts.

Install

Search OpenH3-IR in ComfyUI Manager and install, or manually:

git clone https://github.com/ruashots/ComfyUI-OpenH3-IR.git ComfyUI/custom_nodes/ComfyUI-OpenH3-IR
/path/to/ComfyUI/python -m pip install -r ComfyUI/custom_nodes/ComfyUI-OpenH3-IR/requirements.txt

The second command installs open-h3-ir into the Python ComfyUI runs, so compilation happens in-process - nothing to launch, no port to pick. You'll also need a Setup node wired in (your LLM endpoint plus the five H3 files) and a Media node if anything in your sentence is a @ reference.

When something goes wrong

Read the report. It's built to tell apart failures that look identical: a missing language-model address, an OpenH3-IR install that won't import, an endpoint serving several models. Two things burn people: the LLM needs vision if your job contains visual references (Setup tests this for you), and a mention of a media slot that doesn't exist turns red and refuses the queue before any model call. When in doubt, the report names the slot and the reason.