[Chen Wang]

Hey! I’m trying to recreate the recent famous 90s Yearbook AI trend where the user uploads 8 ~ photos of themselves and the AI generates a whole bunch of photos in the style of 90s yearbook.

See example here: https://imgur.com/a/j2u7idC

Or just google 90s yearbook ai trend.

I’m pretty sure what they’re doing is a Dreambooth model + Controlnet Canny on a base image and use prompts to control the clothing + background.(In some photos, you can find variance in the clothing).

However, what I find really impressive is how good their faces are.

Not only does it really look like the end user, but it also looks really good. I suspect they might be doing a faceswap because on some of the photos I see a little blur around the edges of the face. However, I’ve tried faceswap with ReActor, but the results often don’t look like my target face that much and the skin is often smoothened.

Does anyone have any tips as to how to achieve that?

[Andrew]

They could be doing roop or IP-adapter with inpainting, or both. Perhaps the hardest part is to match the skin color with the face. They must have some kind of global generation like IP-adapter, and followed by swapping the face with roop.

I am curious to try it out in A1111 if you can find a sample dataset of inputs and outputs.

[Author]

Posts