Update(info) about the coloring model and Rife-App (Patreon)
Content
Coloring model:
Ok, for the coloring model, I knew that GAN was really hard to train, but I gave it a shot anyway. It was hard but in the end it generated colored images. The quality was fine for a showcase, but nothing that could really be used in real life.
I did learn a lot, some stuff that I'm using on Rife-App now.
So for now I put Gan aside and started with Diffusion models. I had a feeling that this would be the correct path to take, but I do need to throw everything on the wall, see what sticks.
Currently my Diffusion model trained for about 7% of the total training, but is already showing some progress.
I gotta to say, even at 7% training, is already more creative with the colors than the GAN model.
My ultimate goal is to have 3 models.
1: Sketch->Color
2: Grayscale->Color
3:Live Action photo/video -> Color
Rife-App:
Let me resume why this last step is taking a little more time that I would like:
1) A few weeks ago I did a totally crap model, full of erros and stuff, but it trained very, very well. The best model I had a few weeks ago.
2) So I fixed the wrong stuff on this model and trained again.
3) Surprise, now the model is slight worse?
In the mean time, I found a few new techniques that improve the model overall, it's better to find patters, but still didn't gave the sharp edges of this model a few weeks ago.
So now, I'm trying to merge both the new stuff and whatever worked so well on this older model, to try to make a really good model. It's almost there, each try I get closer to get to great results.