LoRa, which I trained on my own computer, is still working on how to adjust parameters and improve the data set.The current version is 0.1, waiting to be fully perfected.

Sansochan

Retrain based on IL 1.0, and use more consistent tags to improve the training effect.Reclean and create synthetic data to improve quality and better preserve details.It is recommended to use WAI-NSFW-illustrious-SDXL v14.0 output to alleviate some overfitting problems.Since the details of various combinations are not compared using magnified details, the headdress and bow on the sansochan wedding dress may not be effectively eliminated on the back and are easy to generate.It supports some NSFW but not many, and the total number of training samples including original appearance is about 150.When closer to the face or upper body, eye detail is better.

Files with SHA256: c6f3f2f8a9f3f5cf84430d171dbc2283a5d9e40515f295edd44d4da08da2667a