Introduction

MiniMax H3 is gloriously picky — and that is a compliment. Open-weight H3 video wants structured prompts: named sections, shot timings, speaker IDs, and tidy media tags. MiniMax’s rewriting model H3-Context-IR never shipped open-source, so creators were left translating vibes into ritual by hand.

Enter ComfyUI-Fantastic-MiniMaxH3-PromptBuilder (~168★, MIT, Nodes 2.0 compatible, currently around 1.7.2): the sunny, hand-driven replacement. Fillable templates for every mode, live guide checking, and a media loader that keeps your tags honest — so your desk stays golden and your queues stay sane.

What it is (and why H3 smiles)

This is a ComfyUI custom node pack for guided prompt writing plus reference-media handling for open-weight MiniMax H3. Think of it as a storyboard desk that already knows the house style: you fill the fields, it assembles a correctly formatted prompt, and it flags the oopsies before you burn a long render.

Nodes live under conditioning → video_models:

  • Fantastic H3 Prompt Builder — the main editor
  • Fantastic H3 Media Loader — drag-and-drop references with real tag numbers
  • Fantastic H3 Reference Splitter (optional)
  • Fantastic H3 Filename Prefix (optional)

Plus the RefMod stack — Create / Edit / Inspect / Text Encode — so you can save a look, place, or voice once and reuse it like a favorite lipstick shade.

Modes under the sun

Templates for all five H3 lanes: T2VA, I2VA, FL2VA, L2VA, and Reference. Same sunny workflow: build the graph, open Edit prompt…, pick the matching mode, fill, Save to node, queue.

Pro Tip

Wire the Media Loader (or keyframes) before you write. Click-to-insert picture / video / audio tags from thumbnails beats typing picture_1 / video_1 tags by candlelight — and live checking will still catch unused refs, shot-number drama, and dialogue that wandered off script.

Creator highlights (1.6 → 1.7.x)

  • Click-to-insert media tags from thumbnails; hover previews that feel like candy
  • Live checking for shot numbers, cut times, dialogue, and unused refs
  • Built-in 📖 guide (HTML — searchable, phone-friendly)
  • Draft mode — scratchpad while Live is queued; drafts can carry their own media/RefMods
  • Prompt library + media presets + linked prompts
  • Non-destructive trim/crop; pull frames and audio from clips without touching originals
  • RefMods (1.7.x) — save look/voice/place once; Draft from RefMods; subject names like !Ann; voice descriptions that ride with the speaker buttons
  • Delivery / community performance tags (pauses, whispers, laughs) from 1.6.4+
  • Security hardening in 1.6.2 — path containment, CSRF-style POST checks, no ffmpeg shell-out (PyAV only)

For the full RefMods walkthrough, read the upstream guide on GitHub: REFMODS.md.

Install

Needs ComfyUI ≥ 0.30.0 and H3 models (fl2va / ref2va, or a fl2va/ref2va hybrid — hybrids are the warm recommendation for reference quality). Full ComfyUI restart required after install (not just a browser refresh).

Via git:

cd ComfyUI/custom_nodes
git clone https://github.com/Adudeguyman/ComfyUI-Fantastic-MiniMaxH3-PromptBuilder

Or open ComfyUI Manager, search Fantastic H3 Prompt Builder, install, restart.

Confirm by searching the node menu for “MiniMax H3” — the Fantastic nodes should appear under conditioning → video_models.

Conclusion

H3’s structure is a feature, not a scold. This pack turns that structure into a playful editor: templates for every mode, live guide checks, media tags that stay true, Draft mode for the next shot, and RefMods so your favorite face or voice does not evaporate between workflows.

Clone it, restart ComfyUI, and let the sunshine into your H3 graph.

Repo: github.com/Adudeguyman/ComfyUI-Fantastic-MiniMaxH3-PromptBuilder

—Aurelia ♡