fix for SDXL

now supports SDXL

README.md

@@ -39,11 +39,7 @@ Or manually clone this repo into your extensions folder:
 
 ![Screenshot of the slider provided by the extension in UI](/assets/screenshot.png "Does what it says on the box.")
 
-After installing, you can locate the new parameter "Negative Prompt Weight" in the extentions area of txt2img and img2img tabs. 
+After installing, you can find the new parameter "Negative Prompt Weight" in the extentions area of txt2img and img2img tabs. 
-
-## Limitations
-
-Doesn't work with SDXL yet.
 
 ## More Comparisons and Stuff
 

scripts/npw.py

@@ -4,7 +4,8 @@ import gradio as gr
 
 import modules.scripts as scripts
 import modules.shared as shared
-from modules.script_callbacks import CFGDenoiserParams, on_cfg_denoiser, remove_current_script_callbacks
+from modules.script_callbacks import on_cfg_denoiser, remove_current_script_callbacks
+from modules.prompt_parser import SdConditioning
 
 
 class Script(scripts.Script):
@@ -65,7 +66,7 @@ class Script(scripts.Script):
             # print('NPW callback removed')  
 
         if self.weight != 1.0:
-            self.empty_uncond = self.make_empty_uncond() 
+            self.empty_uncond = self.make_empty_uncond(p.width, p.height) 
             on_cfg_denoiser(self.denoiser_callback)
             # print('NPW callback added')
             self.callbacks_added = True    
@@ -79,25 +80,37 @@ class Script(scripts.Script):
     def postprocess(self, p, processed, *args):
         if hasattr(self, 'callbacks_added'):
             remove_current_script_callbacks()
+            delattr(self, 'callbacks_added')
             # print('NPW callback removed in post')
 
     def denoiser_callback(self, params):
+        def concat_and_lerp(empty, tensor, weight):
+            if tensor.shape[1] > empty.shape[1]:
+                num_concatenations = tensor.shape[1] // empty.shape[1]
+                empty_concat = torch.cat([empty] * num_concatenations, dim=1)
+                if tensor.shape[1] == empty_concat.shape[1] + 1:
+                    # assuming it's controlnet's marks(?)
+                    empty_concat = torch.cat([tensor[:, :1, :], empty_concat], dim=1)
+                new_tensor = torch.lerp(empty_concat, tensor, weight)
+            else:
+                new_tensor = torch.lerp(empty, tensor, weight)
+            return new_tensor
+
         uncond = params.text_uncond
+        empty_uncond = self.empty_uncond
+        is_dict = isinstance(uncond, dict)
 
-        if uncond.shape[1] > self.empty_uncond.shape[1]:
+        if is_dict:
-            num_concatenations = uncond.shape[1] // self.empty_uncond.shape[1]
+            uncond, cross = uncond['vector'], uncond['crossattn']
-            empty_uncond_concat = torch.cat([self.empty_uncond] * num_concatenations, dim=1)
+            empty_uncond, empty_cross = empty_uncond['vector'], empty_uncond['crossattn']
-            if uncond.shape[1] == empty_uncond_concat.shape[1] + 1:
+            params.text_uncond['vector'] = concat_and_lerp(empty_uncond, uncond, self.weight)
-                # assuming it's controlnet's marks!
+            params.text_uncond['crossattn'] = concat_and_lerp(empty_cross, cross, self.weight)
-                empty_uncond_concat = torch.cat([uncond[:, :1, :], empty_uncond_concat], dim=1)
-            new_uncond = torch.lerp(empty_uncond_concat, uncond, self.weight)
         else:
-            new_uncond = torch.lerp(self.empty_uncond, uncond, self.weight)
+            params.text_uncond = concat_and_lerp(empty_uncond, uncond, self.weight)
 
-        params.text_uncond = new_uncond
+    def make_empty_uncond(self, w, h):
-
+        prompt = SdConditioning([""], is_negative_prompt=True, width=w, height=h)
-    def make_empty_uncond(self):
+        empty_uncond = shared.sd_model.get_learned_conditioning(prompt)
-        empty_uncond = shared.sd_model.get_learned_conditioning([""])
         return empty_uncond
 
     def print_warning(self, value):