The generation of kagura demuri's clothes seems to be unstable, you need some luck.

It seems difficult for multiplayer lora to handle every detail of each character perfectly.

If I have a high-performance GPU, I will train until I am satisfied. But on colab, I don't want to spend too much computing units.

LoRa trained on civitai is overfitting again. Maybe 330 images are too many.

There seems to be no way to adjust the learning rate when training lora in civitai. Both adafactor and prodigy are adaptive optimizers, which will automatically set both Unet LR and Text Encoder LR to 1 (I only tried prodigy).

amata sora

amata sora,black t-shirt,shirt,vest,red vest,fingerless gloves,belt

kagura demuri

kagura demuri,fur trim,kagura's shirt,bracelet,single glove,dark blue pants

cayenne suzushiro

cayenne suzushiro,jacket,jacket on shoulders,vest,green vest,sheathed,dagger