1 files changed, 11 insertions, 6 deletions

diff --git a/training/functional.py b/training/functional.py
index 62b8260..a9c7a8a 100644
--- a/training/functional.py
+++ b/training/functional.py
@@ -256,6 +256,7 @@ def loss_step(
     text_encoder: CLIPTextModel,
     with_prior_preservation: bool,
     prior_loss_weight: float,
+    low_freq_noise: float,
     seed: int,
     step: int,
     batch: dict[str, Any],
@@ -274,13 +275,15 @@ def loss_step(
         layout=latents.layout,
         device=latents.device,
         generator=generator
-    ) + 0.1 * torch.randn(
-        latents.shape[0], latents.shape[1], 1, 1,
-        dtype=latents.dtype,
-        layout=latents.layout,
-        device=latents.device,
-        generator=generator
     )
+    if low_freq_noise > 0:
+        noise += low_freq_noise * torch.randn(
+            latents.shape[0], latents.shape[1], 1, 1,
+            dtype=latents.dtype,
+            layout=latents.layout,
+            device=latents.device,
+            generator=generator
+        )
     bsz = latents.shape[0]
     # Sample a random timestep for each image
     timesteps = torch.randint(
@@ -553,6 +556,7 @@ def train(
     global_step_offset: int = 0,
     with_prior_preservation: bool = False,
     prior_loss_weight: float = 1.0,
+    low_freq_noise: float = 0.05,
     **kwargs,
 ):
     text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler, extra = strategy.prepare(
@@ -587,6 +591,7 @@ def train(
         text_encoder,
         with_prior_preservation,
         prior_loss_weight,
+        low_freq_noise,
         seed,
     )