I should be studying C++ but since it's Sunday I decided to procrastinate on this new pet project I came up with.
Through David, I found some interesting inspiration to generate cool AI tattoos. Here's an example.
I've contacted the dude who created this dataset to learn more about the configurations he used, if he pre-trained stuff, etc. I've compiled a gist with initial pointers to this.
First, I'd like to understand what's the difference between some random collabs that I found and the original one. Some of them claim to be better to train your own model on top.
After running a test in a Stylegan3+CLIP notebook I reached the conclusion (yes, after one experiment, they take too long!) that the model is the soul of this stuff. I reached out to the dude who created the inspirational thing, maybe he still has this dataset somewhere and would be up to giving it to me. Anyways, I think I'll start to create my own.
Okay, so creating a model seems like a bit crazy. Ideally I'd find a drawing model.
Then, I'd fine tune it with my own stuff that I'll scrape and/or the dude will share with me.
Apparently, all you have to do for that is using the --resume flag. This colab is pretty straight forward.