Upload pipeline.py with huggingface_hub

pipeline.py

@@ -89,6 +89,10 @@ class StableDiffusionT5Pipeline(StableDiffusionPipeline):
         )
 
         if t5_projection is None:
+            print("WARNING: no CLIPTextModelWithProjection found. This may indicate an error")
+            answer=input("Should I auto-generate one? type 'Yes' to proceed")
+            if answer != "Yes":
+                exit(1)
             self.t5_projection = self.create_clipholder().to(vae.device, dtype=vae.dtype)
         else:
             if isinstance(t5_projection, CLIPTextModelWithProjection):
@@ -96,6 +100,17 @@ class StableDiffusionT5Pipeline(StableDiffusionPipeline):
             else:
                 raise TypeError("Error: expected t5_projection to be type CLIPTextModelWithProjection")
 
+        checkval = getattr(self.t5_projection.config, "scaling_factor", None)
+        if not checkval:
+            #0.013 # This is my kinda calculated factor, for norms ~ 1.0
+            #scaling_factor = 0.13025 # This would be the vae scaling factor
+            #scaling_factor = 0.035    # This is a commonly used factor for T5
+            # buuut... to make output stdD similar to CLIP, scaling factor = 1.8
+            # (See check-cache-stdd-t5.py)
+            scaling_factor = 1.8
+            print("INFO: Pipeline setting empty t5 scaling factor to", scaling_factor)
+            self.t5_projection.config.scaling_factor = scaling_factor
+
         # Ensure everything is properly registered for to("cuda")
         # and also for saving the model
         self.register_modules(t5_projection=self.t5_projection)
@@ -137,10 +152,8 @@ class StableDiffusionT5Pipeline(StableDiffusionPipeline):
         **kwargs,
     ):
 
-        # Scaling value computed by analyzing stock output over a few thousand prompts,
-        # and computing stdD 
-        # vs alleged expected value of 0.2
-        scaling_factor = 1.8 # coincidentally, this is about the same as vae.config.scaling_factor
+
+        scaling_factor = self.t5_projection.config.scaling_factor
 
         pos_hidden, neg_hidden = self.encode_prompt_t5(prompt, negative_prompt, device)