Codename "JL" is an artificial character that was created in research lab. The dataset was heavily edited in multiple software and not intended to imitate any person in real.
Lora was created for research and testing purposes.
===========================================================
Beta48 was trained without Text Encoder quantization, and optimizer EPS changed to 1e-8
===========================================================
Both Beta19 and Beta29 increase the precision to fp32, but Beta19 is dedicated for NSFW
===========================================================
Beta15 increases the dimensions to 256.
===========================================================
Beta11 increases the dimensions to 128.
===========================================================
Beta7 increases the dimensions to 64.
===========================================================
Beta4 increases the dimensions to 32, in order to hold more details.
===========================================================
Beta1 was found over trained after many pictures were drawn. Therefore the Beta2 is released with the steps reduced.
===========================================================
jlFemale FLUX Beta1 is the prototype version of LoRA for the flux.1-dev model
jlFemale is another character that created in lab, also for research purposes. The objectives are looking for optimal dataset and training setting in order to produce the best result.
After the introducing of new model FLUX, the research of lora is also expanded to this model. Please feel free to visit other loras for more details and create pictures.
Description
The AI-Toolkit settings are as follows:
Optimizer: AdamW
Mix Precision: fp32
Save Precision: float32
Learning rate: 1e-4
Unet LR: 1e-4
Linear: 256
Linear_alpha: 256
Bucket Size: 1024x1536
Batch Size: 1
transformer_only: false
caption_dropout_rate: 0
ema removed
quantize: false
T5 8bit quantized