This is really good for training niche concepts or vtubers with only 10 pictures
https://github.com/kohya-ss/sd-scripts/pull/1710It prevents Lora from over-fitting / changing the style and composition too much.
examples: i used 13 train images + 124 reg images. first row - with preservation, 2nd - without, 3nd - no lora
(WARNING: concept is inserting a toilet brush in unexpected places)
https://litter.catbox.moe/5819d7.jpghttps://litter.catbox.moe/1m3ota.jpghttps://litter.catbox.moe/epyome.jpg