A workflow for swapping characters using Minimax H3.
A single go, swap 1 to 4 characters.
How to use
Worker nodes
Worker nodes. They do the logic, comparison and prompt creations.
Models
Select your main clip, diffusion model and additional clip for the prompt assembling - I recommend qwen3vl_8b_int8_convrot.
Main settings
Select the image to edit.
Nodes:
Settings: the main settings. You can override image size to a custom size. You can choose not to use fizgig (single latent, about 20 percent speedup, but the more character you use the less). You can and must set the number of characters from 1-4 here. Choose the reference image size. Reinforce relight adds an extre prompt automatically for all subjects to relight them, which usually helps with placing them more naturally, but you are free to turn it off.
Additional overall instruction: you can ask for extra edits here besides the basic character swap. Fixing, pose adjustment or even background change is possible here.
Instruct plus: sometimes you might have to adjust the main instruction prompt to get the subjects you want. I mainly used it for things like "ignore background characters" or "treat the chibi plush as character".
Characters
Set the character reference images here.
If needed you can add extra notes to the character - like "she's a minigirl, downsize her" or "age her up".
Preview instructions
Here you can check if the model got the subject you wanted to as well as determine which character is which. It's usually left-to-right, but it might change, so sometimes it's worth to give it a quick glance.
Result
You can check the assembled final instruction prompt here.
The final image will be saved here as well.
Custom nodes needed:
Fizgig H3 Still: https://github.com/shootthesound/ComfyUI-Fizgig-H3-Still
rgthree node pack: https://github.com/rgthree/rgthree-comfy