Stable Diffusion merging two faces to make one

Jan 30, 2023
17
7
I'm trying to merge two faces of two women I really like. I'd like to make a lora, or even just an image I can reuse in reactor.
I don't want to use the girls real face as that would violate her consent, so I though of merging her with another hot girl i know....creating a whole new face...

I've tried dreambooth but my gtx 1660ti can't handle it.

Any suggestions?
 

osanaiko

Engaged Member
Modder
Jul 4, 2017
2,548
4,634
Are these real world normal people?

If you are talking about celebrities you can put both names in the prompt with and you'll get a mix. The relative strength can be varied with strength adjustment i.e. (scarlett johanssen:0.7),(claudia schiffer:0.5)
 
Last edited:
Nov 26, 2018
46
22
You should probably use a combination of youtube + google than a forum for sex games. Might find what you're looking for quicker. But yes -- you can train to combine 2 faces. From what I know about it (I've never done it but I've watched videos) - you'll need to do the following:

Get about 15 detailed photos in varying poses, clothing, closeups, full body, etc. You're looking for quality over quantity. You'll wan to trim each photo or image so that they are all the same square size for training. You can train in any resolution, but it needs to be square. The higher the resolution, the more time you'll need to invest, but you'll end up with a potentially better Lora that will do well at higher resolutions.

You'll want to create a txt file for each image which describes the postive prompts for that image. You'll want to include a unique identifier to describe the merge photo which you're trying to create. You will need to create a unique and non-common identifier for the Lora. Avoid common names like: "1girl, woman, lady," etc. Think of what you want the lora to be called, and use that as your identifier. Include this unique identifier for each text file for each image in your training folder. Eventually after thousands and thousands of iterations, you should come up with a workable Lora that will have a unique face. Then when you use the Lora, you'll just need to use the unique identifier and you can start creating...

That's what I know about it, though I highly recommend you don't take my word for it, and you look up a few videos.

From what I know, it's always quality over quantity.

I don't know how many iterations you'll need in total to get a workable model, and I don't know how many steps you should use for each. But...best of luck!

Civitai is a good resource to learn more about modeling (and downloading new models and loras).