ComfyUI — MiniMax H3 Prompt Nodes

Seven nodes that assemble spec-compliant MiniMax H3 prompts, so you write shot content and the scaffolding is generated for you.

Specs implemented:

• VIDEO_PROMPT_WRITING_GUIDE_base_en.md — T2VA / I2VA / FL2VA / L2VA
• VIDEO_PROMPT_WRITING_GUIDE_ref_en.md — full-reference (r2v)

---

Install

Unzip so you end up with:

ComfyUI\custom_nodes\comfyui-minimax-h3\__init__.py

Restart ComfyUI. No dependencies — pure stdlib.

On startup the console prints:

[MiniMax H3] wrote ...\comfyui-minimax-h3\web\minimax_h3_buttons.js
[MiniMax H3] serving at /extensions/comfyui-minimax-h3/minimax_h3_buttons.js
[MiniMax H3 Prompt] registered 7 nodes: ...

The web folder is written by the Python on every start, so there is only one file to install and the script can't fall out of sync. If the buttons don't appear, hard-refresh the browser (Ctrl+Shift+R) — the old script gets cached.

All nodes appear under the MiniMax H3 category.

---

The two chains

Base modes — one Shot node per shot:

Shot → Shot → Shot → Assemble → prompt / total_seconds / frames

Full reference (r2v) — Subject chain plus the same Shot chain:

Subject → Subject ─┐
                   ├→ Ref Prompt Builder r2v → prompt / total_seconds / frames
Shot → Shot ───────┘

---

Nodes

MiniMax H3 Shot

cut_verb, seconds, text, and a shots chain socket (leave empty on the first). Cut timestamps are computed cumulatively — set 5 / 2 / 3.5 and you get [Shot 2] At 00:05.0, and [Shot 3] At 00:07.0, automatically. Inserting a node mid-chain renumbers everything downstream.

Two buttons under the text box wrap a highlighted selection in <d>[English] ...</d> or in double quotes for on-screen text.

MiniMax H3 Assemble

Terminates a Shot chain. mode picks T2VA / I2VA / FL2VA / L2VA and the right alignment instruction is prepended. FL2VA and L2VA take their duration and last shot index from the chain, so they can't drift from what you actually wrote.

MiniMax H3 Subject

One subject_definitions line. Two phrasings, chosen by the role:

• content role → <Subject 1> is <description>, whose <role> comes from <reference>.
• frame role → <Picture 1> is the <role> of <reference>, showing <description>.

role_2 / reference_2 handle a subject drawing on two assets — appearance from an image, motion from a video.

MiniMax H3 Ref Prompt Builder r2v

The six sections in fixed spec order. subjects and shots chain inputs override their text boxes when connected. Three ordered task_type dropdowns build the bracketed summary prefix.

MiniMax H3 Multi Shot Builder

Fixed eight-slot alternative to the Shot chain, with one global cut verb.

MiniMax H3 Prompt Builder

Single-body builder — you write the whole description including shot headers.

MiniMax H3 Shot Timer

Durations in (one per line), formatted shot headers out. Useful when typing a body by hand instead of chaining Shot nodes.

---

Notes

24fps is hardcoded. MiniMax H3 is trained at 24fps, so it's a module constant rather than a widget. The frames output snaps to the nearest valid 17n+5 length — the lattice isn't evenly spaced, so a shot can shift by up to ~0.35s.

Cut times use one decimal (00:05.0). The FL2VA/L2VA alignment line keeps two decimals, which the spec requires.

FL2VA bracket style is reproduced verbatim from the official guide, which is internally inconsistent — I2VA and L2VA use <Picture 1> / [Shot 1] while FL2VA uses bare Picture 1 / Shot 1. Set NORMALISE_BRACKETS = True near the top of __init__.py to make them match, if FL2VA misbehaves.