Fix training for V-pred and ztSNR

1) Updates debiased estimation loss function for V-pred.
2) Prevents now-deprecated scaling of loss if ztSNR is enabled.

library/custom_train_functions.py

@@ -96,10 +96,13 @@ def add_v_prediction_like_loss(loss, timesteps, noise_scheduler, v_pred_like_los
     return loss
 
 
-def apply_debiased_estimation(loss, timesteps, noise_scheduler):
+def apply_debiased_estimation(loss, timesteps, noise_scheduler, v_prediction=False):
     snr_t = torch.stack([noise_scheduler.all_snr[t] for t in timesteps])  # batch_size
     snr_t = torch.minimum(snr_t, torch.ones_like(snr_t) * 1000)  # if timestep is 0, snr_t is inf, so limit it to 1000
-    weight = 1 / torch.sqrt(snr_t)
+    if v_prediction:
+        weight = 1 / (snr_t + 1)
+    else:
+        weight = 1 / torch.sqrt(snr_t)
     loss = weight * loss
     return loss
 

library/train_util.py

@@ -3731,6 +3731,11 @@ def verify_training_args(args: argparse.Namespace):
         raise ValueError(
             "scale_v_pred_loss_like_noise_pred can be enabled only with v_parameterization / scale_v_pred_loss_like_noise_predはv_parameterizationが有効なときのみ有効にできます"
         )
+    
+    if args.scale_v_pred_loss_like_noise_pred and args.zero_terminal_snr:
+        raise ValueError(
+            "zero_terminal_snr enabled. scale_v_pred_loss_like_noise_pred will not be used / zero_terminal_snrが有効です。scale_v_pred_loss_like_noise_predは使用されません"
+        )
 
     if args.v_pred_like_loss and args.v_parameterization:
         raise ValueError(