Allow individual models to scale the loss after it is calculated.

extensions_built_in/sd_trainer/SDTrainer.py

@@ -827,6 +827,9 @@ class SDTrainer(BaseSDTrainProcess):
                 loss = torch.nn.functional.mse_loss(pred.float(), target.float(), reduction="none")
             
             loss = loss * local_loss_scale
+            
+            # apply model specific loss scaling
+            loss = self.sd.scale_loss(loss)
                 
             do_weighted_timesteps = False
             if self.sd.is_flow_matching:

toolkit/models/base_model.py

@@ -1602,3 +1602,7 @@ class BaseModel:
     def get_model_to_train(self):
         # called to get model to attach LoRAs to. Can be overridden in child classes
         return self.unet
+    
+    def scale_loss(self, loss):
+        # called to get the loss scaler for the model. Can be overridden in child classes
+        return loss

toolkit/stable_diffusion_model.py

@@ -3161,3 +3161,7 @@ class StableDiffusion:
 
     def get_model_to_train(self):
         return self.unet
+        
+    def scale_loss(self, loss):
+        # called to get the loss scaler for the model. Can be overridden in child classes
+        return loss