This describes my first experiments with this dataset. I will also be tweaking it with adafactor
I was initially playing around with loras, to see which combination of images did best. Previously, I spent a LOT of time experimenting with training and getting nowhere, so I decided to stick with adaptive ones. Sadly, even a 4090 cant do a finetune of SDXL with an adaptive optimizer, which is why I was training Loras.
My results were very mixed. Eventually, I figured out that stable diffusion, while seemingly magic, cannot TRUELY figure out "good" anime style, if you throw a whole bunch of MIXED styles at it. So I decided to drastically change my strategy, and throw out everything that was not strictly in one style.
Once I got down to <200 images images, and had a reasonable lora, I decided to give a full finetune a try, "the hard way" (ie: no adaptive optimizer)
So, I set EMA=CPU (because not enough VRAM to fit in GPU) played with the learning rate a little, and... that was it? Nope!
I was training with 100 epoch, and having onetrainer do an image sample every few epochs. ut, due to disk space, I was only doing saves every 20 or so.
When I looked back at the image samples, I noticed that the one I liked best, was actually around epoch 72. didnt know what e71,72, or 73 looked like... and I didnt even have a save for 72 either!
What I did have, was a save at 70. So I configured OneTrainer to do a new run, starting with my saved model at 70. This time, however, I disabled warmup, and also EMA. I then set OneTrainer to make a save every epoch, and a preview every epoch.
The previews looked kind of like my original series! Thus encouraged, I decided to try out all 10 of the models, just in case. It turns out that I liked epoch 78/100 the best.. so here we are :)