Two-Reference Image Editing — MiniMax H3
A ComfyUI workflow built on the MiniMax H3 model, designed to edit images using two reference images at once. The first image sets the foundation — the scene, object, or character — while the second brings in the desired change: different clothing, an accessory, or a new object to add to the original photo. The model blends both sources carefully, preserving the composition and recognizability of the original.
Great for outfit changes, adding objects, and targeted object-level edits. ⚠️ This workflow does not change pose — the original pose is preserved as-is.
Use MiniMax-H3-Ref2VA-pruned_int8_convrot.safetensors for best result
The workflow has been updated. fixed an error with file names when saving. Don't judge me harshly - this is the first time I've published it.
Description
FAQ
Comments (9)
Can you give an example of how it works? I tried it and I. Nothing happened with the same image was generated.
Published original photo. Prompt: In photograph1, place the mechanism with photograph2.
Make the mechanism small to the middle of the fence height.
@Osa ok lets say your swapping the pose of a couple, how would you do that, couple together in one image but you want to swap them to the pose in the 2nd photo
Maybe I spoke too soon when I said that it’s possible to transfer the pose from the reference. You can add objects and change clothes, but the pose isn’t transferred. However, this is a model for generating videos, not for editing. I’ll try to make it transfer the pose specifically, but it might not be possible.
Prompt: In photograph1, place the mechanism with photograph2.
Make the mechanism small to the middle of the fence height
This is nice but I want to generate an image of different size than <Picture 1>, just the generic one megapixel size, how I disable it in your workflow?
Yes it is possible. You need to add one node to the subgraph, and there you can choose either through sets of formats and megapixels. Or add a manual size indication.
How about person replacing? Like: replace person on image1...&
I can't say it's necessary to try, but in general it should work.