MiniMax H3 Context IR Refiner

Raw H3 prompts get better results when someone polished them first - and MiniMax itself offers a hosted refiner, "H3 Context IR," that takes your rough shot description and turns it into the detailed, structured prompt the model rewards. This node calls that official API. It's the polish layer for people who want the best prompt possible without running a local LLM. The name is honest: it refines, it doesn't generate video - the result is a MINIMAX_H3_PROMPT that then flows into Conditioning like any other prompt.

How it works

You feed it a prompt (from SimplePrompt or Storyboard) and it sends your text - plus any reference media from a package - to MiniMax's Context IR endpoint, along with your instruction and music_style. It gets back a polished prompt and repackages it into the same payload shape Conditioning expects. There's a built-in preview panel that shows the refined result (mode, frame count, duration, ratio, and the final prompt), and it polls for ComfyUI cancellation between API requests so a queue cancel actually stops it.

Two things make this node notable. First, the API key never enters the workflow JSON: it's read from the IR_KEY environment variable at execution time. That's the right call for a node that reaches the network - this is exactly the API-wrapper pattern the community has learned to treat carefully, and env-var keys are the safe shape of it. Second, it's a hosted call: per use, this costs API credits. You're paying MiniMax for the polish, which is a reasonable trade when it's a good refiner.

The inputs that matter

• instruction (required) - the refiner's marching orders. Default: "Make it more cinematic, detailed and temporally clear." Edit to taste.
• music_style (required) - non-diegetic music requirement; empty means N/A.
• prompt (optional) - a MINIMAX_H3_PROMPT from SimplePrompt or Storyboard.
• package (optional) - PACKAGE_DATA so the refiner can see and describe your reference media.
• timeout - seconds per API request, default 300.

Outputs: prompt (the refined MINIMAX_H3_PROMPT, into Conditioning) and text (plain STRING of the polished prompt). The node is marked as an output node because of the preview panel.

Installing it

Pack-wide routine - ComfyUI Manager search "MiniMax H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/xiaolibai-sys/ComfyUI-MiniMaxH3
pip install -r requirements.txt

then restart. Then set the key: export IR_KEY=your_key_here (or set it in your ComfyUI launch environment). No model files - the refiner lives in MiniMax's cloud.

Common issues

• 401 / auth errors. IR_KEY isn't set in the environment ComfyUI runs in. Export it before launching ComfyUI - putting it in the node won't work, by design.
• Slow first call. A hosted refiner round-trip over your reference media takes real seconds. timeout defaults to 300 for a reason.
• Cost surprise. Every execution is a paid API call. Batch your prompt-polish sessions rather than running the workflow repeatedly while tuning unrelated nodes.
• Key in a shared workflow. That's the whole point of the env-var design - the key stays out of the JSON, so sharing workflows stays safe. Keep it that way.