diff --git a/examples/community/lpw_stable_diffusion_xl.py b/examples/community/lpw_stable_diffusion_xl.py
index 7c825c8dbabb..abfbfb5aa1c1 100644
--- a/examples/community/lpw_stable_diffusion_xl.py
+++ b/examples/community/lpw_stable_diffusion_xl.py
@@ -1184,13 +1184,19 @@ def __call__(
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
 
-        # make sure the VAE is in float32 mode, as it overflows in float16
-        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
-            self.upcast_vae()
-            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
-
         if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
         else:
             image = latents
             return StableDiffusionXLPipelineOutput(images=image)
diff --git a/examples/community/stable_diffusion_xl_reference.py b/examples/community/stable_diffusion_xl_reference.py
index 45cd4e0a2e0e..b47c962701b6 100644
--- a/examples/community/stable_diffusion_xl_reference.py
+++ b/examples/community/stable_diffusion_xl_reference.py
@@ -772,13 +772,19 @@ def hacked_UpBlock2D_forward(self, hidden_states, res_hidden_states_tuple, temb=
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
 
-                # make sure the VAE is in float32 mode, as it overflows in float16
-        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
-            self.upcast_vae()
-            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
-
         if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
         else:
             image = latents
             return StableDiffusionXLPipelineOutput(images=image)
diff --git a/src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py b/src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py
index c99f007f9c82..cd2f3daf55b4 100644
--- a/src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py
+++ b/src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py
@@ -1183,13 +1183,19 @@ def __call__(
             self.controlnet.to("cpu")
             torch.cuda.empty_cache()
 
-        # make sure the VAE is in float32 mode, as it overflows in float16
-        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
-            self.upcast_vae()
-            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
-
         if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
         else:
             image = latents
             return StableDiffusionXLPipelineOutput(images=image)
diff --git a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py
index a0cadaaa52d1..4661dd78f215 100644
--- a/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py
+++ b/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py
@@ -751,15 +751,20 @@ def __call__(
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
 
-        # 10. Post-processing
-        # make sure the VAE is in float32 mode, as it overflows in float16
-        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
-            self.upcast_vae()
-            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
-
-        # post-processing
         if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+
             image, has_nsfw_concept, _ = self.run_safety_checker(image, device, prompt_embeds.dtype)
         else:
             image = latents
diff --git a/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py b/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py
index 11e575d68269..2d4ef87bdf79 100644
--- a/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py
+++ b/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py
@@ -870,13 +870,19 @@ def __call__(
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
 
-        # make sure the VAE is in float32 mode, as it overflows in float16
-        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
-            self.upcast_vae()
-            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
-
         if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
         else:
             image = latents
             return StableDiffusionXLPipelineOutput(images=image)
diff --git a/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py b/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py
index ff51f8765e4a..aada52eecbcb 100644
--- a/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py
+++ b/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py
@@ -1027,13 +1027,19 @@ def denoising_value_valid(dnv):
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
 
-        # make sure the VAE is in float32 mode, as it overflows in float16
-        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
-            self.upcast_vae()
-            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
-
         if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
         else:
             image = latents
             return StableDiffusionXLPipelineOutput(images=image)
diff --git a/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py b/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py
index eecbdc7e669e..bc29cecdbdc9 100644
--- a/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py
+++ b/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py
@@ -1333,13 +1333,19 @@ def denoising_value_valid(dnv):
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
 
-        # make sure the VAE is in float32 mode, as it overflows in float16
-        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
-            self.upcast_vae()
-            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
-
         if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
         else:
             return StableDiffusionXLPipelineOutput(images=latents)
 
diff --git a/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py b/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py
index 83345454443f..45c90b0d8d66 100644
--- a/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py
+++ b/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py
@@ -908,13 +908,19 @@ def __call__(
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
 
-        # make sure the VAE is in float32 mode, as it overflows in float16
-        if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
-            self.upcast_vae()
-            latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
-
         if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
         else:
             image = latents
             return StableDiffusionXLPipelineOutput(images=image)