Switched to new bucket system that matched sdxl trained buckets. Fixed requirements. Updated embeddings to work with sdxl. Added method to train lora with an embedding at the trigger. Still testing but works amazingly well from what I can see

toolkit/config_modules.py

@@ -72,6 +72,7 @@ class TrainConfig:
         self.lr = kwargs.get('lr', 1e-6)
         self.unet_lr = kwargs.get('unet_lr', self.lr)
         self.text_encoder_lr = kwargs.get('text_encoder_lr', self.lr)
+        self.embedding_lr = kwargs.get('embedding_lr', self.lr)
         self.optimizer = kwargs.get('optimizer', 'adamw')
         self.optimizer_params = kwargs.get('optimizer_params', {})
         self.lr_scheduler = kwargs.get('lr_scheduler', 'constant')