MiniMax H3 Prompt Builder

The MiniMax H3 Prompt Builder looks like the kind of node that should be phoning home to an API. It's not. This thing makes zero network calls, needs no key, and doesn't even require the H3 model to be installed - it's pure text construction that turns ten messy fields into one structured, section-labeled prompt. That's the whole job, and it's more useful than it sounds, because MiniMax H3 isn't Stable Diffusion.

H3 is the 33B omni-modal video model from the lab behind Hailuo: text, image, video, and audio all live in one context, and it generates stereo sound with the picture instead of bolting an audio pass on afterward. A flat "cinematic shot of a robot in a workshop" prompt is throwing away its best trick. It rewards you for saying what the scene sounds like, how the camera moves, and what actually changes over the clip. This node exists to make you write that richer prompt without hand-rolling the format every time.

How it works

Each field becomes a labeled line, with empty fields dropped:

task_mode: T2VA (text-to-video with audio)
subject: A ceramic robot with a weathered blue glaze
scene: A quiet workshop during a rainstorm
action_over_time: The robot slowly turns toward the window and raises one hand
camera: Slow medium tracking shot
lighting: Warm practical lamps against cool window light
motion_and_continuity: Continuous natural movement with stable identity
overall_soundscape: Soft motor movement, rain ambience, and distant thunder

Two details matter. First, reference_instructions only gets appended when your mode is I2VA, FL2VA, or REF2VA - the modes that actually have input media to describe. Text-only T2VA silently skips it, which is correct behavior, not a bug. Second, the node's second output isn't just a copy of the prompt; it's a validation_report generated by the same structural checks the MiniMax H3 Prompt Validator runs. So you build and get a quick structural once-over in one step.

The inputs that matter

mode is the one you must get right first - T2VA (text-to-video with audio), I2VA (image-to-video with audio), FL2VA (first-and-last-frame), or REF2VA (reference-guided). Pick the family that matches the workflow you're feeding.

After that, subject, scene, and action are your core three - the "what", the "where", and the "what changes over time". camera, lighting, and motion fill out the visual language. The two everyone forgets are dialogue and soundscape; on an audio-native model they're not garnish, they're half the point.

The prompt output wires into the text input of your H3 conditioning workflow. The validation_report goes wherever you'd read a string - a text viewer node, or just the console.

Install

Open ComfyUI Manager, search for MiniMax H3 Workflow Toolkit, and install it. Or, if you'd rather:

cd ComfyUI/custom_nodes
git clone https://github.com/jayseanbrambila-del666/ComfyUI-MiniMax3-H3-Toolkit

Restart ComfyUI and search the node library for "MiniMax H3". That's the whole install - this pack is deliberately dependency-free. No pip install, no model downloads, nothing but a requirements.txt that says "No third-party Python dependencies."

Common gotchas

The big one: this node produces text, not video. If you connect it and hit Run expecting a clip, nothing renders, because the toolkit ships no model files at all. The actual H3 weights (roughly 42.5GB) and the conditioning setup are a separate install, which the author's guide at minimax3.org covers. Think of this pack as the prompt side of the pipeline.

Worth knowing before you go further: the H3 weights themselves ship under the MiniMax H3 Community License, which excludes the US, EU, UK, and South Korea - if you're in one of those regions, only the hosted Hailuo API path is licensed. And don't mistake this for an official MiniMax node: the pack's README is explicit that it's an independent community utility, not affiliated with MiniMax or Comfy Org.