008
The laugh the retouch erased
What happens when I ask the AI to improve a face that was already fine?
Part 01
An automatic step
When I like an image, I do two things. First I make it twice as big with an upscaler called AnimeSharp. Then I run a face retouch: a module called FaceDetailer finds the face and paints it again with more detail. To paint it, the module needs a text that says what the face looks like. I used the same text for every image.
Krea 2 · AnimeSharp 2x · FaceDetailer, face and eyes · seed 812207 · 2048×25604 attempts
Four versions of the same face
In this image she laughs with her mouth open and her eyes closed. My usual text said “a beautiful adult woman's face” and nothing about laughing. The retouch did what I asked: a beautiful face, and a serious one. Then I changed only the text and described the laugh. I tried it with two seeds.
- F1Starting point
The enlarged image, with no retouch. She laughs with her eyes closed.
AnimeSharp 2x · no retouch - F2It didn't work
Sharper, but the laugh is gone. It opened her eyes and closed her mouth.
FaceDetailer · face 0.35 · eyes 0.20 · my usual text - F3It worked
I described the laugh and it came back. But it is not the same laugh: now her eyes are open.
Same settings · text that describes the laugh · seed 812207 - F4It worked
Another seed, same text. She laughs too. In all three retouched versions her earrings changed.
Same settings · text that describes the laugh · seed 812217
What I learned
The retouch does not improve the face that is there. It paints the face the text describes.
If the expression matters, I describe it in the text of each image. The laugh came back in 2 out of 2 tries.
Even then it changes things I did not ask for. Here it opened her eyes and swapped her earrings.
In other series the same step changed the eye color and added grain to the skin. So now I check every face before and after.
If the face is already fine, I leave it alone. Sometimes I retouch only the eyes. Sometimes nothing.
Attempts: F1
The next experiment
The “no” that isn't there
How do I say “no” to a model that has no negative?
OpenFollow the lab
Every experiment goes up as it happens: the picked images, the clips in motion and the batches that didn't work.
Write to me on Instagram
