diff --git a/library/train_util.py b/library/train_util.py
index 7d311827..e68444a0 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -1927,7 +1927,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         parser.add_argument(
             "--prior_loss_weight", type=float, default=1.0, help="loss weight for regularization images / 正則化画像のlossの重み"
         )
-
+    parser.add_argument("--min_snr_gamma", type=float, default=0, help="gamma for reducing the weight of high loss timesteps. Lower numbers have stronger effect. 5 is recommended by paper.")
 
 def verify_training_args(args: argparse.Namespace):
     if args.v_parameterization and not args.v2:
diff --git a/train_network.py b/train_network.py
index 7f910df4..5cb08f15 100644
--- a/train_network.py
+++ b/train_network.py
@@ -548,6 +548,16 @@ def train(args):
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
                 loss = loss * loss_weights
+                gamma = args.min_snr_gamma
+                if gamma:
+                  sigma = torch.sub(noisy_latents, latents) #find noise as applied
+                  zeros = torch.zeros_like(sigma) 
+                  alpha_mean_sq = torch.nn.functional.mse_loss(latents.float(), zeros.float(), reduction="none").mean([1, 2, 3]) #trick to get Mean Square
+                  sigma_mean_sq = torch.nn.functional.mse_loss(sigma.float(), zeros.float(), reduction="none").mean([1, 2, 3]) #trick to get Mean Square
+                  snr = torch.div(alpha_mean_sq,sigma_mean_sq) #Signal to Noise Ratio = ratio of Mean Squares
+                  gamma_over_snr = torch.div(torch.ones_like(snr)*gamma,snr)
+                  snr_weight = torch.minimum(gamma_over_snr,torch.ones_like(gamma_over_snr)).float() #from paper
+                  loss = loss * snr_weight
 
                 loss = loss.mean()  # 平均なのでbatch_sizeで割る必要なし