Fixed a bug in the prompt embeddings.

The original prompt embeddings will cause the pipeline to crash in the case the negative prompt was longer than the prompt. Implemented a fix suggested in huggingface/diffusers#2136.

src/diffuser_tools/text2img.py

@@ -156,57 +156,43 @@ def get_prompt_embeddings(
         """Prompt embeddings to overcome CLIP 77 token limit.
         https://github.com/huggingface/diffusers/issues/2136
         """
+        count_prompt = len(self.prompt.split(split_character))
+        count_negative_prompt = len(self.negative_prompt.split(split_character))
 
         max_length = self.pipe.tokenizer.model_max_length
 
-        # Simple method of checking if the prompt is longer than the negative
-        # prompt - split the input strings using `split_character`.
-        count_prompt = len(self.prompt.split(split_character))
-        count_negative_prompt = len(self.negative_prompt.split(split_character))
+        input_ids = self.pipe.tokenizer(
+            self.prompt, return_tensors = "pt", truncation = False
+        ).input_ids.to(self.device)
+        negative_ids = self.pipe.tokenizer(
+            self.negative_prompt, return_tensors = "pt", truncation = False
+        ).input_ids.to(self.device)
 
-        # If prompt is longer than negative prompt.
-        if count_prompt >= count_negative_prompt:
-            input_ids = self.pipe.tokenizer(
-                self.prompt, return_tensors = "pt", truncation = False,
-            ).input_ids.to(self.device)
+        if input_ids.shape[-1] >= negative_ids.shape[-1]:
             shape_max_length = input_ids.shape[-1]
             negative_ids = self.pipe.tokenizer(
-                self.negative_prompt,
-                truncation = False,
-                padding = "max_length",
-                max_length = shape_max_length,
-                return_tensors = "pt",
+                self.negative_prompt, return_tensors = "pt", truncation = False, 
+                padding = "max_length", max_length = shape_max_length
             ).input_ids.to(self.device)
-        # If negative prompt is longer than prompt.
         else:
-            negative_ids = self.pipe.tokenizer(
-                self.negative_prompt, return_tensors = "pt", truncation = False,
-            ).input_ids.to(self.device)
             shape_max_length = negative_ids.shape[-1]
             input_ids = self.pipe.tokenizer(
-                self.prompt,
-                return_tensors = "pt",
-                truncation = False,
-                padding = "max_length",
-                max_length = shape_max_length,
+                self.prompt, return_tensors = "pt", truncation = False, 
+                padding = "max_length", max_length = shape_max_length
             ).input_ids.to(self.device)
 
-        # Concatenate the individual prompt embeddings.
         concat_embeds = []
         neg_embeds = []
         for i in range(0, shape_max_length, max_length):
-            concat_embeds.append(
-                self.pipe.text_encoder(input_ids[:, i: i + max_length])[0]
-            )
-            neg_embeds.append(
-                self.pipe.text_encoder(negative_ids[:, i: i + max_length])[0]
-            )
+            concat_embeds.append(self.pipe.text_encoder(input_ids[:, i: i + max_length])[0])
+            neg_embeds.append(self.pipe.text_encoder(negative_ids[:, i: i + max_length])[0])
 
         self.prompt_embeddings = torch.cat(concat_embeds, dim = 1)
         self.negative_prompt_embeddings = torch.cat(neg_embeds, dim = 1)
 
         if return_embeddings is True:
-            return torch.cat(concat_embeds, dim = 1), torch.cat(neg_embeds, dim = 1)
+            return self.prompt_embeddings, self.negative_prompt_embeddings
+
 
     # %%
     def run_pipe(
@@ -258,4 +244,4 @@ def run_pipe(
         if verbose is True:
             sys.stdout.write("{:.2f}s.\n".format(time_elapsed));
 
-        return imgs[0]
+        return imgs[0]