Fixed the overfitting problem in previous versions.
Even larger training dataset (6,617 images)
Auto tagged with wd-v1-4-moat-tagger-v2 with zero pruning.
Used flip augmentation and color augmentation.
Used multi resolution noise during training.
Obtained from LoCon extraction after fine-tuning the entire model.
Downscaled the LoCon dimension to 128.
Fixed the overfitting problem in previous versions.
Even larger training dataset (6,617 images)
Auto tagged with wd-v1-4-moat-tagger-v2 with zero pruning.
Used flip augmentation and color augmentation.
Used multi resolution noise during training.
Obtained from LoCon extraction after fine-tuning the entire model.
Downscaled the LoCon dimension to 128.