A tribute to the late Dean "Hutch" Hutchinson, rest in peace.
I tried to improve the quality of my traning, and thus tried to tweak my training parameters before finding my original parameters were very good for creating a flexible style and not overtrain/overfit my small dataset. So in the end, I just went with full fp16 training, no half vae and added flip augmentation. Seems a tad better imho.
No trigger word. Recommended negatives : big eyes, very big eyes, signature, text, english text











