MiniMax-H3 Prompt Presets

If you've stared at a MiniMax-H3 text input and wondered what it actually wants, this is the node that answers without asking anything of your GPU. MiniMax-H3 doesn't take a sentence. It wants a structured, three-field audio-video description - a shot-by-shot integrated_multimodal_description with [Shot 1], [Shot 2] markers and cut times, an overall_soundscape, a non_diegetic_music line - because that's the shape the model was trained to read. Writing that by hand is the fiddly part of every H3 workflow, and whole packs exist just to generate it from a one-line idea. MiniMax-H3 Prompt Presets skips the generating entirely: it hands you someone else's finished prompt, one of a thousand that ship inside the pack.

It's the laziest node in the MiniMax-H3-Prompt-Rewriter-ComfyUI pack, and it knows it. Where its sibling rewriter nodes load a 27B language model and spend tens of seconds turning "a red fox walks through a snowy forest at dawn" into the labeled format, this one reads a prompt off disk and passes it on. A run takes single-digit milliseconds. No model is loaded, nothing is downloaded, no VRAM moves - the whole library of prompts, tags and a 256-pixel frame per clip is about 6 MB bundled with the pack.

The prompts are ostris's work from the minimax_h3_1k dataset, carried in with his permission. His prompts are effectively MIT; the clips they describe are MiniMax-H3 output, which means the frames carry the model's licence and you shouldn't use them to train anything.

How you actually use it

Press Pick a preset and a browser opens over the thousand, narrowed by twenty shooting styles, twenty subjects, the shape of the frame, the shot count and whether anyone speaks, with a search box over the words themselves. Each row shows the frame of the clip that prompt was written for - click it and the clip plays, the only thing that touches the network. Pick one and the whole prompt, all three labeled fields, appears right on the node's face before you've run anything. The run then just hands it out.

Where it goes is your call, and that's a wire rather than a widget:

• prompt straight into the H3 generator's text input - it's a complete T2VA prompt, used as it stands.
• prompt into one of the pack's writer nodes' own prompt input - it becomes the starting point for a rewrite, which is handy when a preset is close but the shot pacing or the soundscape isn't yours.

The preset input is just a number - the "Pick a preset" button writes it, but typing one works too. Because that number is stored in the workflow, a graph you share resolves to the same prompt on somebody else's machine. The other outputs split the text apart when you want pieces: integrated_multimodal_description, overall_soundscape and non_diegetic_music are the three fields on their own, while seconds, width and height describe the clip it was written for - every one is about five seconds - and source carries its number, both watch addresses and the credit, handy for a text preview when the workflow travels. Honestly, 95% of the time you wire prompt and nothing else.

Installing it

Same as the rest of the pack. In ComfyUI Manager search "MiniMax-H3 Prompt Rewriter", or:

cd ComfyUI/custom_nodes
git clone https://github.com/pytraveler/MiniMax-H3-Prompt-Rewriter-ComfyUI

then restart. One honest gotcha: the pack's requirements.txt pulls in transformers, peft, accelerate and bitsandbytes for the rewriter nodes, and installing the pack means installing those - even though this node alone needs none of them and you could run it on a potato. If 50 GB of models daunts you, this node plus the writer templates are the entry point.

The trap worth knowing

These are T2VA prompts for a five-second clip, and the shot times inside the text are written against that. Nothing stops you handing one to a reference task or a sixty-second render, and nothing will complain - but the text describes neither. Treat them as a format sample as much as a finished product: a great way to learn what a good H3 prompt looks like, and a fast start you'll often want to rewrite anyway. Save to the library files a copy into your own prompt sets, editable and re-checked as you type, if one earns a second use.