Egami Style (NoobAI-XL v-pred)

Expanded a little bit the dataset and did some proper preprocessing to try to remove all the compression artifacts that were present. The training images were ai-upscaled 4x and then downscaled back to a 1024x1024 pixel count. Also changed the training parameters a little bit.

Some training details:

dataset_count = 52

optimizer = CAME

learning_rate = 0.0003

learning_rate_scheduler = cosine

lora_rank = 32

lora_alpha = 1

epochs = 12

steps = 624

batch_size = 1