MiniMax H3 Prompt Validator

H3's prompt format is fiddly enough that you will, at some point, hand it a prompt that looks right and comes back with audio desync or dialogue the model invented. The MiniMax H3 Prompt Validator is the node that catches those before they cost you a render: it checks any prompt against H3's documented audiovisual structure - strictly, deterministically, and with no LLM in the loop. It's the bouncer at the club door. It checks your ID; it does not vouch for your personality.

What it checks

The validator runs the prompt through the same contract parser the enhancer nodes use after enhancement. For the mode you select, it verifies the section blocks are present and correctly ordered - the 6-block Ref2VA anatomy (subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music) or the 3-block T2VA/I2VA/FL2VA/L2VA structure. It checks that quoted dialogue lives inside <d>[Language] ...</d> tags, that the language tag matches, that references like <Picture 1> or <Audio 1> are declared before use, that speaker IDs don't dangle, and that the prompt is free of leftover authoring shorthand.

That last one matters more than it sounds. The pack lets you type delivery marks like 😠 or 🤫 next to a line as shorthand, and it resolves them into prose. A stray bracket that survives to the final prompt used to surface only as mysterious "invented dialogue" - the validator now names the leaked mark directly instead of blaming the wrong cause. It also refuses a prompt containing the emoji palette, which is a neat guarantee that nothing raw reached the model.

The honest caveat is printed in the node's own description: structural validity does not guarantee generation quality. A perfectly formatted prompt can still be boring, or wrong, or logline-nonsense. This node checks grammar, not taste. What it gives you is the confidence that when a render is bad, the prompt format wasn't the reason.

Inputs that matter

• prompt - the prompt to check (from your own pen or from an enhancer's enhanced_prompt output).
• mode - the contract to validate against; auto infers it.
• duration_seconds - used to sanity-check shot timestamps and frame-count math against the 17×n+5 grid.
• source_prompt - your original request. This is what makes verbatim-dialogue checking possible: the validator can flag dialogue the LLM invented that wasn't in your source, which is the pack's defense against the classic subject-drift failure of enhancers.

The rest of the inputs mirror the enhancer's policy widgets (reference context, media manifest, creative latitude, multishot locks, and so on) - set the same values you used to generate the prompt, or the validator will judge it against the wrong contract.

Outputs

prompt (pass-through, so you can daisy-chain it), valid (boolean), and validation_report - a JSON rundown of errors and warnings. It's an output node too, so the report shows up in the UI even mid-workflow.

Install

You know the drill by now - it's in the same pack:

cd ComfyUI/custom_nodes
git clone https://github.com/hyukudan/ComfyUI-MiniMax-H3-Prompt-Enhancer.git

Restart ComfyUI. Zero dependencies, zero API calls, works offline.

Where it fits

The natural slot is between an enhancer and the H3 sampler: enhancer enhanced_prompt → validator → sampler. If you built your own pipeline with the Guide Builder, the validator is your quality gate - feed it the LLM's reply and it tells you whether the model actually followed the spec. The author's framing is that validation plus the automatic repair loop (in the enhancer nodes) is what makes prompts "production-grade"; the standalone validator is the same gate for people running their own writer.