diff --git a/nerfstudio/model_components/renderers.py b/nerfstudio/model_components/renderers.py
index 1f95de95f..6a8e2a5f9 100644
--- a/nerfstudio/model_components/renderers.py
+++ b/nerfstudio/model_components/renderers.py
@@ -37,7 +37,7 @@ from torchtyping import TensorType
 from typing_extensions import Literal
 
 from nerfstudio.cameras.rays import RaySamples
-from nerfstudio.utils.math import components_from_spherical_harmonics
+from nerfstudio.utils.math import components_from_spherical_harmonics, safe_normalize
 
 BACKGROUND_COLOR_OVERRIDE: Optional[TensorType[3]] = None
 
@@ -328,9 +328,18 @@ class NormalsRenderer(nn.Module):
         cls,
         normals: TensorType["bs":..., "num_samples", 3],
         weights: TensorType["bs":..., "num_samples", 1],
+        normalize: bool = True,
     ) -> TensorType["bs":..., 3]:
-        """Calculate normals along the ray."""
+        """Calculate normals along the ray.
+
+        Args:
+            normals: Normals for each sample.
+            weights: Weights of each sample.
+            normalize: Normalize normals.
+        """
         n = torch.sum(weights * normals, dim=-2)
+        if normalize:
+            n = safe_normalize(n)
         return n
 
 
diff --git a/nerfstudio/models/nerfacto.py b/nerfstudio/models/nerfacto.py
index c996e20b3..91b044884 100644
--- a/nerfstudio/models/nerfacto.py
+++ b/nerfstudio/models/nerfacto.py
@@ -250,8 +250,10 @@ class NerfactoModel(Model):
         }
 
         if self.config.predict_normals:
-            outputs["normals"] = self.renderer_normals(normals=field_outputs[FieldHeadNames.NORMALS], weights=weights)
-            outputs["pred_normals"] = self.renderer_normals(field_outputs[FieldHeadNames.PRED_NORMALS], weights=weights)
+            normals = self.renderer_normals(normals=field_outputs[FieldHeadNames.NORMALS], weights=weights)
+            pred_normals = self.renderer_normals(field_outputs[FieldHeadNames.PRED_NORMALS], weights=weights)
+            outputs["normals"] = (normals + 1) / 2
+            outputs["pred_normals"] = (pred_normals + 1) / 2
 
         # These use a lot of GPU memory, so we avoid storing them for eval.
         if self.training: