It's finished! I have images for all of my Dante data set of V2! It was rather hard given that I needed to hook up a ton of full bodies that had transparent backgrounds to an image underneath and then give some attempt at making a shadow that made sense over all that 3D terrain. BluePencil XL was heavily used for this process making a ton of backgrounds. I had said the new data set would expand to make more data and I wasn't wrong. In total, I ended up with 228 images of Dante from the base set of 38 images. Almost all of them new and designed around a consistent color palate and anatomy. About the only thing that varies is the number of tails stripes. There is potential to get that wrong given the AI might figure it out from 36 of the 228 images in the set, but otherwise everything is perfect!
I also balanced the set, as you'll notice that 228 is 2 + 2 + 8 = 12 which is divisible by 3. That is to say that for every solid colored background (alternating and balancing those colors a bit, too) there are 2 complicated backgrounds. This should dramatically help the AI to avoid making "boring backgrounds" while BluePencil XL is employed so that the backgrounds I used in this process were of the highest quality and looked like a great anime picture.
The balancing method is also useful so that no particular picture gets more exposure than the others, reducing the chances of overtraining the AI. The entire process is meticulous and requires a ton of effort, often feeling like industrial scale art. For instance, I needed 768x768 sized images for training 1.5 models, but my art was all in 1024x1024 size so that I could explore how this reacts to an XL scale model in the future. Using Gimp, I set up batch updates that resized all the images and went off for a snack while my computer chomped away, creating a second folder of images. It's really something else to see and build.
I am now off to tag all these images, which is going to be a beast. I should be helped out because I have prompts for all my characters to describe their situation, but I intend to reduce all these down below 75 tokens, something I didn't do for previous experiments. This time I will also be using class images and a significantly higher batch size along with a couple of other LoRA tweaks.
I also plan on training a Dreambooth, as I now have the computer power to pull it off. The exciting part of the game is almost here! :D :D :D
| Viewed: |
13 times |
| Added: |
4 months, 3 weeks ago
12 Jan 2024 03:53 CET
|
|
|
