Author: hlky

src/diffusers/pipelines/controlnet/pipeline_controlnet_union_inpaint_sd_xl.py

@@ -1577,6 +1577,9 @@ def denoising_value_valid(dnv):
 
         original_size = original_size or (height, width)
         target_size = target_size or (height, width)
+        for _image in control_image:
+            if isinstance(_image, torch.Tensor):
+                original_size = original_size or _image.shape[-2:]
 
         # 10. Prepare added time ids & embeddings
         add_text_embeds = pooled_prompt_embeds

src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl.py

@@ -1290,7 +1290,11 @@ def __call__(
             )
 
         # 7.2 Prepare added time ids & embeddings
+        original_size = original_size or (height, width)
         target_size = target_size or (height, width)
+        for _image in control_image:
+            if isinstance(_image, torch.Tensor):
+                original_size = original_size or _image.shape[-2:]
         add_text_embeds = pooled_prompt_embeds
         if self.text_encoder_2 is None:
             text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])

src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl_img2img.py

@@ -1430,7 +1430,11 @@ def __call__(
             )
 
         # 7.2 Prepare added time ids & embeddings
+        original_size = original_size or (height, width)
         target_size = target_size or (height, width)
+        for _image in control_image:
+            if isinstance(_image, torch.Tensor):
+                original_size = original_size or _image.shape[-2:]
 
         if negative_original_size is None:
             negative_original_size = original_size