Update qwenimage/pipeline_qwen_image_edit.py

qwenimage/pipeline_qwen_image_edit.py

@@ -744,11 +744,19 @@ class QwenImageEditPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
             self._attention_kwargs = {}
 
         txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist() if prompt_embeds_mask is not None else None
-        negative_txt_seq_lens = (
-            negative_prompt_embeds_mask.sum(dim=1).tolist() if negative_prompt_embeds_mask is not None else None
-        )
 
         image_rotary_emb = self.transformer.pos_embed(img_shapes, txt_seq_lens, device=latents.device)
+        if do_true_cfg:
+            negative_txt_seq_lens = (
+                negative_prompt_embeds_mask.sum(dim=1).tolist()
+                if negative_prompt_embeds_mask is not None
+                else None
+            )
+            uncond_image_rotary_emb = self.transformer.pos_embed(
+                img_shapes, negative_txt_seq_lens, device=latents.device
+            )
+        else:
+            uncond_image_rotary_emb = None
 
         # 6. Denoising loop
         self.scheduler.set_begin_index(0)
@@ -786,7 +794,7 @@ class QwenImageEditPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
                             guidance=guidance,
                             encoder_hidden_states_mask=negative_prompt_embeds_mask,
                             encoder_hidden_states=negative_prompt_embeds,
-                            image_rotary_emb=image_rotary_emb,
+                            image_rotary_emb=uncond_image_rotary_emb,
                             attention_kwargs=self.attention_kwargs,
                             return_dict=False,
                         )[0]
@@ -846,4 +854,4 @@ class QwenImageEditPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
         if not return_dict:
             return (image,)
 
-        return QwenImagePipelineOutput(images=image)
+        return QwenImagePipelineOutput(images=image)