Nice attempt, thank you for reporting it.
I done more test, and i've discovered something interesting using my settings.
I've trained 2 same TI, one on sd1.5 checkpoint and one i my last merged.
On sd1.5 getting chromatic aberration is fast, on TI trained on it max cfg is 7, if you going up you get chromatic aberration .
On my models, i'm able to do cfg 15, with same parameter and same image for train.
The conclusion is the base model do a lot on difference when training.
I trained on "train embedding tab", with [fileword],[name] and do good description of every image for train
Also i started to train an "oily" helper, just for experimenting with train. And i made image my-self on daz3d.
And this is "boom" on train. let me explain:
I've rendered 20 images at 768x768 in different position of a oiled woman,with transparent background. then added a description for every image.
when you train, the image you use is resized at 512x512. but during resize, the alpha channel gone. this is bad.
So i checked "do not resize image" and " Use PNG alpha channel as loss weight ". and really the train go better, and result is good from start.
Also used 16 vector, with 8 word for initialization.
Actually i'm training another at different step, then doing some test before maybe release it.
At the end, i thing using image with alpha is good, so you put only subject in description.
This is my impression, i need to inspect more.
I've to test training face with alpha, this is next test i want to do.
