MiniMax H3 US Config (CRT)

MiniMax H3 doesn't take a prompt like a normal image model. It takes a structured brief - the visual description, the soundscape, the music - and the MiniMax H3 US Config (CRT) node is where all of that lives, along with your seed, your keyframes and every reference you want REF2VA to look at. It bundles the whole per-run setup into one config_pipe and hands it to the Unified Sampler. Think of it as the "what do you want" half of the pipeline; the sampler is the "go do it" half.

This is the pattern CRT uses across its video packs (the LTX2.5 Unified Sampler set works the same way), and once you've used it, the payoff is clear: you iterate on a prompt or swap a reference frame without touching a dozen sampler widgets.

The prompt structure

This is the part people get wrong first. H3's official prompt format uses labeled sections, and the node's prompt box follows it exactly:

• integrated_multimodal_description: - shots, motion, camera, what's on screen.
• overall_soundscape: - ambient sound, dialogue, SFX.
• non_diegetic_music: - optional, the score.

In REF2VA mode you address your media inline with tags: <Picture 1>, <Video 2>, <Audio 3>. If a reference isn't tagged in the prompt, the model has no instruction to use it - the tags are the hook.

Inputs that matter

The required pair is prompt (multiline, and it supports dynamic prompts) and seed (with the usual randomize-on-generate control). Then the optional inputs split by mode:

• I2V: First Frame (I2V) and Last Frame (I2V) - your two keyframes.
• REF2VA: up to nine Ref Image sockets, three Ref Video sockets (an IMAGE batch at 24 fps - the author recommends 48+ frames), three Ref Video Audio sockets paired with those videos, and three standalone Ref Audio sockets.

There are also two overrides, both defaulting to 0 (meaning "keep the sampler's setting"): Frames (override) - snap a specific frame count, ceil-snapped to H3's 17n+5 grid - and MegaPixels (override) - force a resolution instead of the sampler's megapixels_target. Handy when one workflow needs to be shorter or higher-res than the default.

Output: a single config_pipe that goes straight into the Unified Sampler's config_pipe socket.

Install

Part of CRT-Nodes by PGCRT. ComfyUI Manager → search CRT-Nodes → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

Restart ComfyUI. On pack updates, a node showing red or NaN gets fixed by right-click → Fix node (recreate).

Gotchas

Keep REF2VA inputs inside the official limits - nine images, three videos, three standalone audios, twelve files max mixed - because that's the cap the model was trained with, and stuffing more in just gets ignored. Ref videos should be 24 fps with decent length. And the standing H3 caveat: the weights are licensed for local use only outside the US, EU, UK and South Korea, so check the territory before you lean on this workflow.