DAINAPP
patreon
Update(info) about the coloring model and Rife-App
🕑 Added 2022-08-12 17:15:30 +0000 UTCColoring 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.
Comments
GRisk
Thanks! It's already working great, but just like in this examples, there is some places it refuse to color, i'm just trying to fix this before releasing.
Trevor
Need to say that is really impressive.
