>>72221259hard, but simpler way:
>image-to-image AI generation, where you provide an input image and ask the model to produce a new output based on the input>however, the model essentially wants to "overwrite" the input image, so it's easy to completely erase the original and end up with something entirely unrelated>you need to very, very gently adjust the strength of the model until you find a happy middle ground between the input image and the one the model is trying to produce>be prepared to make dozens or a hundred plus images trying to strike the balance>(this may sound like a lot of manual effort, but you can also just run script to make a hundred images with 1% changes in the slider)easy, but complicated way:
>there's a thing called a "ControlNet" that forces the model to obey "visual" instructions or guidelines when producing an output>QR code ControlNets are really good at producing subtle images inside other images>use a QR code controlnet with your "hidden" image>produces much more consistent results (still need to produce a dozen or so images, but it's more like fine tweaking than not-working-at-all)>but learning how controlnets work is going a bit deep into the weeds...Like, be prepared to churn out a lot of crap before getting something that looks nice.