MiniMax H3 Reference Sheet

Every H3 project starts with the same chore: loading the reference images and clips, re-describing the character, re-picking the location. The Reference Sheet is the pack's way of making that a one-time job. Save a person, object, place, style, voice, or whole project collection once, then load it later and pick the media from an embedded gallery instead of re-wiring the graph.

It's the hub node of the "Reference Sheets" family, and it works with both halves of the pack - the legacy role nodes and the newer Plan v2 references.

How it works

A sheet is a self-contained folder living outside the repo:

ComfyUI/user/default/minimax_h3/reference_sheets/
└── <sheet-name>--<short-id>/
    ├── manifest.json
    ├── images/
    └── audio/

Connect ordinary Load Image / Load Audio nodes, choose Create new, give it a name and description, queue once. The pack saves connected images as PNG and audio as WAV, and the manifest records SHA-256 checksums - so a sheet survives its original workflow and input filenames disappearing. Set MINIMAX_H3_REFERENCE_SHEETS_DIR before launch to point the library elsewhere.

What you actually set

• operation - Load existing (read-only), Create new, or Update existing. Update preserves saved media and appends new non-duplicate connections; it's locked behind confirm_update, and it never overwrites. Create never overwrites another sheet either.
• saved_sheet - pick from the dropdown; the embedded gallery refreshes without typed paths.
• sheet_name, sheet_description, tags - the reusable metadata. Describe the thing, not the workflow: "the red ceramic robot," "neon market at night."
• selected_image_key / selected_audio_key - managed automatically by the gallery. You don't type asset keys.
• audio_start_seconds / audio_duration_seconds - a non-destructive trim window for the audio output. Slice a 2–15 second segment out of a longer saved clip without touching the stored file.

The optional image_1..4 and audio_1..3 inputs take ordinary IMAGE/AUDIO tensors. An image batch expands into separate saved references up to H3's nine-image limit; audio tops out at three clips.

Outputs

reference_sheet carries the selection for the legacy role nodes. selected_image is the currently picked image - connect it to a Plan v2 Image Reference and choose its relationship there. selected_audio is the trimmed segment. sheet_report tells you where things were saved and what's selected.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMax-H3-Guide

Restart ComfyUI and look under MiniMax H3/Reference Sheets. No Python dependencies; you only need ComfyUI with native MiniMax H3 support. The pack ships a small web component that powers the gallery, so keep the extension's web assets intact.

Where people get burned

The audio trim is the subtle one: audio_duration_seconds counts toward H3's 15-second cumulative reference-audio limit, and if less audio remains than you asked for, the segment just ends early. And a sheet is reusable - remember that. The description you write on Create is what later reads as the reference's meaning, so a lazy "some robot" haunts every future project that loads it. Duplicate the Reference Sheet node if you want two different segments of one saved clip in the same workflow.