NowaJoestar Style SDXL

UNET only trained model on small dataset. Due to presence of old images in dataset resembles averaged over time nowa's style. Feel free to try getting distortions with following resolutions:
bucket 0: resolution (672, 1024), count: 2 (3:2)

bucket 1: resolution (768, 1024), count: 2 (3:4)

bucket 2: resolution (768, 1344), count: 109 (16:9)

bucket 3: resolution (832, 1184), count: 23 (5:4)

bucket 4: resolution (832, 1248), count: 100 (7:5)

bucket 5: resolution (864, 1152), count: 112 (3:4)

bucket 6: resolution (896, 1120), count: 30 (5:4)

bucket 7: resolution (1024, 672), count: 1 (3:2)

bucket 8: resolution (1024, 768), count: 8 (3:4)

bucket 9: resolution (1024, 1024), count: 189 (1:1)

bucket 10: resolution (1120, 896), count: 38

bucket 11: resolution (1152, 864), count: 96

bucket 12: resolution (1184, 832), count: 29

bucket 13: resolution (1248, 832), count: 100

bucket 14: resolution (1344, 768), count: 102
Personally don't like shadowing on initial release, plan to add noise offset.