revert changes in pipelines.stable_diffusion_xl folder, can seperate PR later

Author: yiyixuxu

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py

@@ -912,6 +912,12 @@ def prepare_latents(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
+
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -925,11 +931,6 @@ def prepare_latents(
             init_latents = image
 
         else:
-            latents_mean = latents_std = None
-            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl_img2img.py

@@ -867,6 +867,12 @@ def prepare_latents(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
+
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -880,11 +886,6 @@ def prepare_latents(
             init_latents = image
 
         else:
-            latents_mean = latents_std = None
-            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/kolors/pipeline_kolors_img2img.py

@@ -609,6 +609,12 @@ def prepare_latents(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
+
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -622,11 +628,6 @@ def prepare_latents(
             init_latents = image
 
         else:
-            latents_mean = latents_std = None
-            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py

@@ -917,6 +917,12 @@ def prepare_latents(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
+
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -930,11 +936,6 @@ def prepare_latents(
             init_latents = image
 
         else:
-            latents_mean = latents_std = None
-            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/pag/pipeline_pag_sd_xl_img2img.py

@@ -707,6 +707,12 @@ def prepare_latents(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
+
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -720,11 +726,6 @@ def prepare_latents(
             init_latents = image
 
         else:
-            latents_mean = latents_std = None
-            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -695,6 +695,12 @@ def prepare_latents(
                 f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
             )
 
+            latents_mean = latents_std = None
+            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
+                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
+            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
+                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
+
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
@@ -708,11 +714,6 @@ def prepare_latents(
             init_latents = image
 
         else:
-            latents_mean = latents_std = None
-            if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None:
-                latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1)
-            if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None:
-                latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1)
             # make sure the VAE is in float32 mode, as it overflows in float16
             if self.vae.config.force_upcast:
                 image = image.float()