Doka ドカ (Artist Style) [PonyXL]

875

New Version applys the style way better with a hell of lot more details.

I changed the way i will train my loras now.

it seems using repeats to increase the steps for training works way better than having high epoch numbers.

Dataset is still 41 images.

Settings i used now are:
Batch 4, repeats 5, steps 2450 (9800 without batch), epoch 50, lora rank 32

Please go an appreciate the original artist DOKA ドカ

promts are: ama rei, by doka, ... (not sure if you can leave them, didn't test)

Weight 1.0 seems to be working best. Feel free to test out different weights

every second picture in my example is without lora for comparison :)

정렬