Pin Diffusers Fix current issue (#31)

- Pine Diffusers (955231c40caa12e0795d8ac1c761b8094b84761a)
- Update requirements.txt (5fdf9b706d02177f95d382b57c7a08189e820ef0)
- Update README.md (8bc46f5d846d4efc9450bdeb8809f5dfa34baadb)
- Update app.py (6a391fbb26799060b04a6966390d1f1b55561665)


Co-authored-by: Radamés Ajna <[email protected]>

README.md

@@ -4,7 +4,7 @@ emoji: 🙃
 colorFrom: red
 colorTo: indigo
 sdk: gradio
-sdk_version: 3.44.4
+sdk_version: 4.13.0
 app_file: app.py
 pinned: false
 license: openrail

app.py

@@ -150,37 +150,46 @@ def call(
 
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = pipe.prepare_extra_step_kwargs(generator, eta)
-        
+
         # 7. Prepare added time ids & embeddings
         add_text_embeds = pooled_prompt_embeds
         add_text2_embeds = pooled_prompt2_embeds
-        
-        # Default dtype if prompt_embeds or prompt2_embeds are None
-        default_dtype = torch.float32
-        
-        # Check and set dtype for add_time_ids
-        dtype_for_add_time_ids = prompt_embeds.dtype if prompt_embeds is not None else default_dtype
+
         add_time_ids = pipe._get_add_time_ids(
-            original_size, crops_coords_top_left, target_size, dtype=dtype_for_add_time_ids
+            original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
         )
-        
-        # Check and set dtype for add_time2_ids
-        dtype_for_add_time2_ids = prompt2_embeds.dtype if prompt2_embeds is not None else default_dtype
         add_time2_ids = pipe._get_add_time_ids(
-            original_size, crops_coords_top_left, target_size, dtype=dtype_for_add_time2_ids
+            original_size, crops_coords_top_left, target_size, dtype=prompt2_embeds.dtype
         )
-        
+
         if negative_original_size is not None and negative_target_size is not None:
             negative_add_time_ids = pipe._get_add_time_ids(
                 negative_original_size,
                 negative_crops_coords_top_left,
                 negative_target_size,
-                dtype=dtype_for_add_time_ids  # Use the same default dtype for negative prompts
+                dtype=prompt_embeds.dtype,
             )
         else:
             negative_add_time_ids = add_time_ids
             negative_add_time2_ids = add_time2_ids
 
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+            prompt2_embeds = torch.cat([negative_prompt2_embeds, prompt2_embeds], dim=0)
+            add_text2_embeds = torch.cat([negative_pooled_prompt2_embeds, add_text2_embeds], dim=0)
+            add_time2_ids = torch.cat([negative_add_time2_ids, add_time2_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        prompt2_embeds = prompt2_embeds.to(device)
+        add_text2_embeds = add_text2_embeds.to(device)
+        add_time2_ids = add_time2_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
         # 8. Denoising loop
         num_warmup_steps = max(len(timesteps) - num_inference_steps * pipe.scheduler.order, 0)
 

requirements.txt

@@ -1,8 +1,8 @@
-diffusers==v0.21.0
+diffusers==0.21.4
 transformers
 accelerate
 torch
 xformers
 Pillow
 filelock
-gradio==3.44.4
+gradio==4.13.0