>>43616875>>43616934You can see here on my royal icing lora even with it going through a 180 something dataset 5 times over it still didnt catch on in the first 1-3 epochs. I used 1e-5 LR. Epoch 4 is maybe perfect but I went with 5 anyway. Having a diverse dataset helps in my case because it hasnt learnt any biases of any characters. Except its maybe hard to get rid of the black panties. Nothing a little emphasis cant fix.
Nice wawa
Danke schön.
>>43619170Ive done it on 11 images LOL.