>>8654808
I went to the dataset again, removed some variations cutting it down to 14 images, added an additional close-up crop of the face as well. Did a full manual tagging pass again, adding additional matching wing tags (demon wings, bat wings, mutliple wings, etc) since it was bleeding through.
Ran prodigy to get an initial good starting learning rate by looking at the tensorboard logs, then switched back to AdamW8bit.
Did a couple of test bakes, tweaking the learning rate for both Unet and TE.
Eventually ended up using this config:
https://files.catbox.moe/7id47n.json