SoulfulCat
The LoRA is trained with arts from wlop (a famous artist) to prove the feasibility of lora training on flux.
It is just a test lora produced by custom training scripts of self-distillation, despite of which the quality of flux model is stilled harmed because of too many training steps -- Training is done in 8k steps of batch size 16 and resolution 512 on an A100 GPU. (There is a bug in version one leads to model collapse and it is produced by overfitting, the bug is solved in v2.)
Lora weight in 0.5~1.0 is suggested.
Popularity
Info
Version v2.0: 1 File
About this version: v2.0
Fix training bug of using guidance scale = 3.5 in training, which leads to model collapse. The version 1.0 utilize model self-distillatation to fix that collapse, but that leads to bad quality. In this version the problem is fixed.
2 Versions
Go ahead and upload yours!
