There are no specific plans for updates.

recommended negative: ?????, cum

v3.0 has increased the weight of active tags to preserve features while maintaining existing picture styles.

Previously, weights of 0.6 to 0.8 were recommended,

v3.0 recommends 0.2 to 0.35.

V1.0: Simple Implementation

V2.0: Increasing Datasets

V3.0: Increase the number of training while increasing the dataset and slightly overfitting the main tag, reducing the overall weight.

V4.0: Remove ambiguous or unstable datasets, change the learning method to Locon, and change the tagging method from individual tags to captions to highlight the main object while reducing conflicts with existing tags.

-V4.1: Edit caption

-V4.2: Additional learning in V4.1 after slightly increasing data

-V5.0: The learning resolution has been increased from 896x896 to 1024x1024.

To be honest, I should have started with 1024x1024 from the beginning, but at the time, I wasn't sure because I was testing it quickly.