krea2-v1:
I suggest using an fp8 or bf16 model as the base model. Due to my local memory limitations, I used an int4 model to demonstrate the output effect. However, LoRa was trained on the original raw model, so I suggest using it on a higher precision model. Please show your images and provide relevant suggestions on the model.
ANIMA:
v2 has updated the training set, added some recent new characters and skins, and added some maps with richer environments. I trained this version three times. The art style of the first two times was always unsatisfactory, so the number of training steps has been increased a lot this time. If you feel that the fingers occasionally break, you can appropriately reduce the model strength.
Illustrious:
v2 Notes: Since the training set of version v1 mainly consisted of character illustrations, the composition diversity was relatively limited. Version v2 enriches the training set.
Version 2 is recommended to have a weight of 0.8
This is the art style of the game Brown Dust 2.
Recommended weight 0.6-0.8.