This Lora is not good enough as I would like it to be. I ran over 25 tests with different settings and dataset adjustments, but I couldn't get anything better.
It took so many tries most likely because of the limited dataset, so I used also screenshots from Koikatsu to train this Lora.
LoRA is designed for use with AutismMix SDXL or other mix
Model: ponyDiffusionV6XL_v6.safetensors
Dataset size: 48
Clip skip: 1
dim32, alpha16
unet_lr=0.0001
te_lr = 0.00005
Constant, AdamW
min_snr_gamma: 7.0
My tips for generating on SDXL for AUTOMATIC1111
Resolution:
1024 x 1024
1152 x 896 or 896 x 1152
1216 x 832 or 832 x 1216
Don't use Hires. fix. Send the picture to img2img, set CFG 0.15-0.25, then select “SD upscale” in the script, you also need to select 4x-AnimeSharp (find out how to install it) or what you like best from upscale. All other settings are default. Click generate.
For me 12gb VRAM is enough for this using --xformers