MiniMax H3 Prompt Builder (Nukun)

If you've ever hand-written a MiniMax H3 prompt, you know the drill: a wall of prose where Scene, Character, Action, Camera, Visual Style and Audio all blur together, and half the queue gets rejected for missing a section. This node exists so that doesn't happen. It assembles one structured H3 video prompt from up to six fixed sections, in the exact order the model expects, and optionally mixes in sampled vocabulary phrases so you can batch out variations without touching the prompt text.

MiniMax is a closed video model - you reach it through its own frontend or an API node, not through a local sampler - so what this node gives you is the prompt side of that workflow. It's a text generator that feeds the thing that actually renders. That's a fine division of labor, and honestly the most common failure in video prompting is upstream of the model anyway.

How it works

Each of the six sections gets a fixed multiline text box (scene_text, character_text, action_text, …) plus a vocab file and two sampling knobs. The *_amount default is 0, which means "no sampling, text only" - you get a clean structured prompt the moment you type something. Raise an amount and the node pulls that many phrases from the section's vocab file using a deterministic shuffle bag, so the same seed-index always gives the same words and incrementing the *_word_index cursor walks through non-overlapping blocks before reshuffling. Empty sections are omitted from the output entirely.

The spoken_dialogue field is the clever bit: type one line and it gets quoted into both Action ("the character speaks: …") and Audio, and the Audio block adds your dialogue_language, dialogue_voice and dialogue_delivery plus a "No other dialogue." instruction. So the model knows who's talking, in what language, with what voice and tone, and that nothing else should be spoken. For a character-driven clip that's most of the work done.

The pack ships six bundled vocab files (resources/minimax_h3_scenes.csv and friends, 80 phrases each) so the defaults work out of the box - scene and audio files in particular are curated to be coherent packages rather than random tag salad.

The outputs you'll actually wire

• prompt - the complete blank-line-separated H3 block, ready to paste or feed downstream.
• scene, character, action, camera, visual_style, audio - each section's finished body without the header, which matters if you want to refine individual sections later.

The natural next stop in this pack is Ollama Video Prompt Refiner (Nukun): its six inputs map 1:1 onto these section outputs, so you can go from randomized structured prompt to a polished, harmonized one in a single hop. You can also just paste prompt into whatever MiniMax client you already use.

Installing it

Same story as every node in this pack:

cd ComfyUI/custom_nodes
git clone https://github.com/OnekoSL/Nukun_ComfyUI_Nodes.git

then restart ComfyUI. Or search "Nukun" in ComfyUI Manager and hit install. No model downloads, no API keys - this node doesn't call anything, it just formats text. The only dependency worth knowing: scipy and PyWavelets are in requirements.txt for the pack's sampler nodes, so the first install may pull those.

Where people get stuck

The trap is the *_amount defaults. Because they're all 0, a beginner who just wants "some random" cranks them to 10 and gets a vocabulary soup that overrides their careful fixed text. The intended pattern is fixed text as the backbone, a small amount (1–3) as seasoning, and 0 when you want the section hand-written. Also: each section's word_index is independent, so incrementing them all at different rates gives you combinatorial variation - which is a feature, but it means "increment the seed" only moves one section at a time. Set the ones you care about on control-after-generate and leave the rest alone.