Trigger words:1girl, hair flower, jewelry, flower, hair ornament, large breasts, Croriin, white hair,red eyes,earrings,bride,wedding dress,thighhighs,necklace,white thighhighs,white gloves,pattern,Grimm Liberen,Croriin

If you don't want the monster to come out, enter monster in Negative

After posting, I don't know why the trigger words were cut off... I have no idea how to fix it...

After some research, I found that a learning rate of about 6e-5 can handle details well, but this lora is learned using 8e-5 Unet and 1e-5 TE.

However, I forgot to train the necklace... I didn't keep this detail well. I originally wanted to train Croriin's flower girl, but because the accessories on the top of the head were not stable after multiple processing, I didn't release lora.

Using a variety of pictures and training multiple iterations, the synthetic data produces a relatively smooth intermediate state, which should be better fitting. At the same time, using comic black and white content makes overfitting difficult and limits the capacity reasonably.

However, the synthetic data has not been carefully checked, and there may be slight high-frequency noise that reduces the quality, so it is not recommended to use a weight of 1 or even more. Please reduce the weight to 0.8~0.9 appropriately.

The training data pollutes the clothing patterns of the original data. Even through tag suppression and decomposition, the pollution will still spread, causing the concept of flowers to be over-generalized and affect the surrounding.

__________

The trained base model uses IL 1.0. Please recommend using the derived WAI-NSFW-illustrious-SDXL v14.0 for better results.

Even with contrastive training, there's no effective guarantee that the earrings will be removed; there's a high probability of "apple-shaped" earrings appearing.

Since most samples, even if marked as wet, will always affect the clothes and body, full of water, due to the main sample's wet appearance.

______________

Relatively free to change body shape and some age appearance, and having many ?????? behaviors is allowed, but although hundreds of samples are as comprehensive as possible, it may still not be enough.