fix for #7365, prevent pipelines from overriding provided prompt embeds

examples/community/lpw_stable_diffusion_xl.py

@@ -827,7 +827,8 @@ def encode_prompt(
                 )
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 prompt_embeds = prompt_embeds.hidden_states[-2]
 
                 prompt_embeds_list.append(prompt_embeds)

examples/community/pipeline_demofusion_sdxl.py

@@ -290,7 +290,8 @@ def encode_prompt(
                 )
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 prompt_embeds = prompt_embeds.hidden_states[-2]
 
                 prompt_embeds_list.append(prompt_embeds)

examples/community/pipeline_sdxl_style_aligned.py

@@ -628,7 +628,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter.py

@@ -359,7 +359,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py

@@ -507,7 +507,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

examples/community/pipeline_stable_diffusion_xl_differential_img2img.py

@@ -394,7 +394,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

examples/community/pipeline_stable_diffusion_xl_ipex.py

@@ -390,7 +390,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/animatediff/pipeline_animatediff_sdxl.py

@@ -438,7 +438,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py

@@ -406,7 +406,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py

@@ -415,7 +415,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py

@@ -408,7 +408,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs_sd_xl.py

@@ -336,7 +336,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py

@@ -321,7 +321,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

@@ -409,7 +409,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -430,7 +430,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py

@@ -534,7 +534,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py

@@ -333,7 +333,8 @@ def encode_prompt(
                 )
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 prompt_embeds = prompt_embeds.hidden_states[-2]
 
                 prompt_embeds_list.append(prompt_embeds)

src/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py

@@ -422,7 +422,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py

@@ -705,7 +705,8 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                pooled_prompt_embeds = prompt_embeds[0] if pooled_prompt_embeds is None else pooled_prompt_embeds
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else: