diff --git a/src/diffusers/models/resnet.py b/src/diffusers/models/resnet.py
index b9718e67f279..68a5341d513f 100644
--- a/src/diffusers/models/resnet.py
+++ b/src/diffusers/models/resnet.py
@@ -40,6 +40,13 @@ def forward(self, hidden_states, output_size=None):
         if self.use_conv_transpose:
             return self.conv(hidden_states)
 
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch
+        # https://github.com/pytorch/pytorch/issues/86679
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
         # if `output_size` is passed we force the interpolation output
         # size and do not make use of `scale_factor=2`
         if output_size is None:
@@ -47,6 +54,10 @@ def forward(self, hidden_states, output_size=None):
         else:
             hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
 
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
         # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
         if self.use_conv:
             if self.name == "conv":
diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
index 00e72de6551a..bd1aade0bc01 100644
--- a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
+++ b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
@@ -327,7 +327,9 @@ def __call__(
         image = self.vae.decode(latents).sample
 
         image = (image / 2 + 0.5).clamp(0, 1)
-        image = image.cpu().permute(0, 2, 3, 1).numpy()
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
 
         safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(self.device)
         image, has_nsfw_concept = self.safety_checker(
diff --git a/src/diffusers/pipelines/stable_diffusion/safety_checker.py b/src/diffusers/pipelines/stable_diffusion/safety_checker.py
index 773a7d4b2107..eedd74e88dfb 100644
--- a/src/diffusers/pipelines/stable_diffusion/safety_checker.py
+++ b/src/diffusers/pipelines/stable_diffusion/safety_checker.py
@@ -36,8 +36,9 @@ def forward(self, clip_input, images):
         pooled_output = self.vision_model(clip_input)[1]  # pooled_output
         image_embeds = self.visual_projection(pooled_output)
 
-        special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().numpy()
-        cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().numpy()
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().float().numpy()
+        cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().float().numpy()
 
         result = []
         batch_size = image_embeds.shape[0]