Used bmaltais's implementation of kohya_ss LoRA training scripts.
This version is resumed from 1.0 with 768x768 images. It will handle usual 512x768 (portrait) generations well, similar to 1.0, 768x512 (landscape) images may result in missing limbs or other weirdness. 768x768 generations are decent.
It is a little overtrained, so for the base SD 1.5 model, I recommend setting the strength around 0.8, and other models, between 0.3 - 0.7.
Demo images for 1.1 were generated with Realistic Vision 2.0!
Total Images: 116 total images, about half high resolution
Regularization Images: 300 total "???? woman" regularization images from ProGamerGov's dataset at huggingface.
Trigger word: legspread a photo of a woman ?????????????????? and showing off her pussy
Training Steps: ~10000 in 1.1, ~17000 total including 1.0
Unet Training Rate: 5.0e-05 (0.00005)
Text Encoder Training Rate: 5.0e-05 (0.00005)
