MiniMax H3 Prompt Composer

So you've seen what a MiniMax H3 video prompt actually looks like - subject_definitions:, [Shot 1] headers with timestamps, <Subject 1> and <Picture 2> tags threaded through every line, dialogue wrapped in <d>[English] "..."</d>. It's a screenplay, not a prompt, and hand-writing one from a blank box is a great way to earn a syntax error three seconds into a 15-second render. This node is 1038lab's answer: a manual composer that knows the H3 format so you can write in it instead of fighting it.

Straight talk about what it is and isn't. MiniMax H3 Prompt Composer doesn't generate anything - no LLM call, no API key needed, no model weights, nothing. It's a text editor dressed as a node. The whole pack is about authoring prompts for MiniMax H3, the open-weights 33B omni video model that ComfyUI supported from day zero (minimax-h3.md covers the model and its famously geofenced community licence). The composer is the hand-authoring lane in a suite whose other lane - H3_Promptor - has an LLM write the screenplay for you.

How it works

The backend is almost insultingly simple: execute() reads whatever text you've typed and hands it back as a string. It returns float("nan") from IS_CHANGED, which is a deliberate trick - the node re-fires on every queue run so the downstream sampler always gets your latest edit. All the actual product lives in the node's custom front end:

• Picking a mode loads the official-style template scaffold for that task (T2VA, I2VA, FL2VA, Ref2VA, V2VA, L2VA, A2V, or blank), so your starter matches what you're actually trying to generate.
• Type @, <, or [ anywhere and a floating autocomplete menu offers the tags H3 expects: <Picture N>, <Subject N>, [Shot N], the section headers like integrated_multimodal_description: and overall_soundscape:, and dialogue markers.
• Syntax highlighting colour-codes sections, tags and timestamps so you can see at a glance if a [Shot 2] block is missing its closing timecode.
• A ✨ Refine button opens a modal that sends your selection, a shot, a subject, or the whole prompt to a configured LLM for polishing. That's the only part that touches a provider - and it's optional.

The inputs that matter

The schema is refreshingly short. mode is the one you set: pick the task template you're scaffolding (it defaults to T2VA). The other input is _composer_prompt, which the tooltip warns you not to connect - the leading underscore means internal storage, where the node persists your text across workflow save/reload.

The single output is PROMPT, a plain STRING: "the composed MiniMax H3 prompt text ready for downstream video generation." Wire it into whatever actually runs H3 - ComfyUI's native H3 sampler, the API path, or a third-party "AIO" node. One gotcha that trips people up: this node is text-only, so <Picture 1> tags in your prose are placeholders. The actual reference image still has to be wired into the sampler separately, and your shot count needs to match the duration you set there.

Install

ComfyUI Manager (search "MiniMax H3-Promptor"), or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/1038lab/ComfyUI-MiniMax-H3-Promptor
pip install -r requirements.txt

Then restart ComfyUI. The dependency list is just requests - no heavy downloads, which is the whole point of a text-writing pack. If you want the Refine button to work, configure an LLM provider under the gear-icon Settings → MiniMax H3 (Ollama and LM Studio are enabled by default; cloud keys are your own).

When to reach for it

Reach for the composer when you want deterministic control and zero API spend - you know the shot list you want and you don't need a director LLM reinterpreting it. Where people get burned: they forget the tags are only half the wiring, or they set mode to I2VA but never attach the first-frame image downstream. Mind the reference media and this is the fastest way to get a clean, schema-safe H3 prompt on the canvas.