Improved by minor edits to the training, smaller batch size and slightly less training. Also removed noise offset as it could easily burn at weight 1 when using multiple LoRAs.
While you could reduce the weight of the model you risk losing details, so it wouldn't be an optimal solution.
Improved by minor edits to the training, smaller batch size and slightly less training. Also removed noise offset as it could easily burn at weight 1 when using multiple LoRAs.
While you could reduce the weight of the model you risk losing details, so it wouldn't be an optimal solution.