RefMod baker

248

Updated: Nov 7, 2026

character

Download

1 variant available

Config Other

KREA 2 - Text2image + RefMod Bake on1WF.json

37.93 KB

Verified: 7 days ago

Download (37.93 KB)

This checkpoint includes a config file, download and place it along side the checkpoint.

Details

Type

Workflows

Stats

134

Reviews

Positive

(7)

Published

Nov 7, 2026

Base Model

MiniMax H3

Hash

AutoV2

About this version

How to use this workflow

1. Make your pictures

On the group switch, turn on Text2Image. The baking group turns off automatically.
Write your image prompt and choose your settings.
Press Run.
Your pictures save here: D:\AI\ComfyUI\ComfyUI\output\Krea2\
Repeat until you have the character pictures you want.

2. Choose pictures for the RefMod

Open that output folder in File Explorer.
Create a folder named after your character.
Copy only your good pictures into it. Include different angles, with a clear, consistent face.
Keep rejected pictures and other characters out. The baker reads every picture in the selected folder, up to 32.

3. Bake the RefMod

On the group switch, turn on Bake RefMods here. Text2Image turns off automatically.
In the folder loader, enter your character’s picture-folder path.
In the baker, choose a new, unique name for the RefMod. It currently says notSidney; reusing an existing name risks replacing that file.
Leave the H3 baking VAE as: minimaxh3\minimax_h3_video_vae_int8_convrot.safetensors
Press Run.
Your baked file saves here: D:\AI\models\refmods\

4. Use it in Continuum

Open your H3 Continuum workflow.
Select the new file in Load H3 RefMods.
Keep the RefMod Bridge connected to the Continuum Sampler’s model input.

The RefMod is for MiniMax H3 video—not Krea image generation. To make more pictures, switch back to Text2Image. Press Ctrl+S whenever you want to save your workflow changes.

The MiniMax H3 Community License Agreement’s grant excludes the European Union, the United Kingdom, the Republic of Korea and the United States of America.

MiniMax H3

update 1.1 added text2image gen on same WF with swtich

++Help keep a character on-model” is accurate; it improves consistency without training a LoRA, but doesn’t guarantee an exact face match.++

This workflow turns a folder of reference images of one character into a reusable H3 RefMod. Load that file into your MiniMax H3 Continuum video workflow to help keep the character’s face and appearance consistent across scenes. It supports realistic or anime references; it doesn’t generate the movie itself.

For MY use, the face is the main goal, yes—but a RefMod can also carry hair, body shape, clothing, and visual style from its reference pictures. It isn’t a face-only lock. but I have not used it

Right now, I’ve only used it for faces. It’s a simpler way to help keep a character on-model in short videos without training a LoRA.

t’s a saved reference-conditioning file, not a trained checkpoint or LoRA. The baker encodes your reference images for H3 to use without retraining the underlying model.