ComfyUI RefMod tutorial: MiniMax H3 character consistency
Learn to create a RefMod from eight character images and use it in ComfyUI with MiniMax H3. The creator reports testing on a 6 GB VRAM GPU.
Player not loading? Watch on YouTube
This ComfyUI tutorial builds a character reference file for MiniMax H3 image generation without traditional LoRA training. The workflow starts with one portrait, generates eight variations, then combines those images into a compact RefMod file. A final generation group loads the file and uses prompts to place the character in different scenes.
The reference preparation stage uses Qwen Image 2.1; the speaker also suggests Flux as an alternative. Users set the image folder and match the maximum item count to the number of references. After saving the RefMod, they refresh the loader and select the character. The speaker says a second RefMod can support a second character.
For local AI hardware, the creator reports testing with 6 GB VRAM and 16 GB RAM. The transcript describes roughly six minutes per image at eight steps and 1.8 megapixels, while the description lists 1 megapixel. These are reported results for the creator's setup, and generation time depends on the PC.
The examples include anatomy and eye errors. The speaker suggests increasing steps or resolution before adjusting reference strength. Installation covers custom nodes, the RefMod text encoder, and Fizgig H3 Still for image encoding and decoding. Users can install missing nodes through the manager, then refresh or restart ComfyUI.
Design Blog
