MiniMax H3 Prompt Validator

Every prompt tool on the market promises to make your generations better. The MiniMax H3 Prompt Validator refuses to make that promise, and that honesty is exactly why you'd use it. It doesn't score quality, it can't predict what H3 will render, and it says so out loud in every report it emits: "This validator checks structure only; it does not predict output quality." If you're coming from the era of "prompt enhancer" nodes that were really just an LLM with a thesaurus, this is the opposite product.

H3 is a 33B omni-modal video model with native audio - it takes text, image, video, and audio as one context. The failure modes people actually hit with it aren't "my prompt wasn't poetic enough," they're structural: no mention of what moves over time, no camera direction, no sound intention, no explanation of what a reference image is for. That's what this validator checks, and it checks it with deterministic keyword matching, not vibes.

How it works

You hand it a mode (T2VA, I2VA, FL2VA, or REF2VA) and the prompt text, and it runs a handful of cheap heuristic checks:

• a prompt under 80 characters gets flagged as too short for a multimodal video task
• no temporal action words ("action", "moves", "walks", "turns", "over time") becomes an issue
• no camera or framing language becomes a suggestion
• no audio language (sound, audio, dialogue, music, ambience) becomes a suggestion
• in I2VA/FL2VA/REF2VA, it insists you actually explain how the input media is used
• FL2VA wants both a first-frame state and a last-frame state described
• REF2VA wants reference roles named with consistent numbering

Issues push status to REVIEW; suggestions alone still give you PASS. The status output is the one-line verdict you skim, and the report output is the full breakdown - every issue, every suggestion, plus the disclaimer.

The input that actually matters

Only two inputs exist here: mode and prompt. Get mode right, because the reference checks are mode-specific - validating a REF2VA prompt as T2VA skips the whole reference-guidance layer, and a report that never mentions your <Picture 1> references isn't giving you the check you asked for.

A natural pairing is to run it on the prompt output of the MiniMax H3 Prompt Builder, which actually produces the builder's validation_report as its second output - same checks, no extra wiring.

Install

Open ComfyUI Manager and search for MiniMax H3 Workflow Toolkit, or clone it straight in:

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

Restart ComfyUI and look for "MiniMax H3" in the node library. There are no third-party dependencies and no model files, so there's nothing else to fetch.

Common gotchas

Set your expectations, or you'll be annoyed. This is a heuristic checker, not an LLM reading for intent - a genuinely great prompt that happens to never say the word "camera" or "sound" will come back REVIEW with suggestions, and that's the tool working as designed. Treat the report as a checklist of what a complete H3 audiovisual prompt should cover, not as a grade.

And the recurring trap with this whole pack: it ships no H3 model. It's text utilities only - you still need the actual weights and conditioning workflow to generate anything, and the H3 Community License geofences the weights out of the US, EU, UK, and South Korea. None of that is this node's problem, but it's worth knowing before you wire everything up and wonder why nothing renders.