Recommended weight 0.65-0.95. Interesting results with more weight, but more artifacts. Sometimes there are problems with generating hair flaps and white sandals, solved by increasing the weights and reroll. My first lora, so it's not ideal at all. It was noticed that sometimes one image from a dataset can be used too often. I haven't tried it with other loras, but I don't think the result will be bad. Triggers: saya, saya no uta, hair flaps. Maybe in the future I'll add images from the visual novel for the official artstyle. Trained on Anything v4.5 --------------------------------------------------------------------------------------------------- V2! In this version I adjusted captions via wd captions, which is why I strongly recommend to use v2 and not v1. I added support for visual novel style via official artstyle, official art, game cg tags. Also try nsfw (included in dataset), eldritch abomination, wings.

Files with SHA256: beeaf72a2fae638bcef2ad44783ff4b6c08be1e93b6780b69f73e1a8172e222b