Modified s_dict generation

app.py

@@ -368,8 +368,9 @@ with blocks:
                         vid_button  = gr.Button("Generate Video")
                         loop_styles = gr.inputs.Checkbox(default=True, label="Loop video back to the initial style?")
                     with gr.Row():
-                        gr.Markdown("Warning: Videos generation requires the synthesis of hundreds of frames and is expected to take several minutes.")
-                        gr.Markdown("To reduce queue times, we significantly reduced the number of video frames. Using more than 3 styles will further reduce the frames per style, leading to quicker transitions. For better control, we reccomend cloning the gradio app, adjusting `num_alphas` in `generate_videos`, and running the code locally.")
+                        with gr.Column():
+                            gr.Markdown("Warning: Videos generation requires the synthesis of hundreds of frames and is expected to take several minutes.")
+                            gr.Markdown("To reduce queue times, we significantly reduced the number of video frames. Using more than 3 styles will further reduce the frames per style, leading to quicker transitions. For better control, we reccomend cloning the gradio app, adjusting `num_alphas` in `generate_videos`, and running the code locally.")
                 with gr.Column():
                     vid_output = gr.outputs.Video(label="Output Video")
 

model/sg2_model.py

@@ -526,7 +526,36 @@ class Generator(nn.Module):
         if not input_is_latent:
             styles = [self.style(s) for s in styles]
 
-        s_codes = [{layer: layer(s) for layer in self.modulation_layers} for s in styles] * len(styles)
+        s_codes = {# const block
+                   self.modulation_layers[0]:  self.modulation_layers[0](styles[0]),
+                   self.modulation_layers[1]:  self.modulation_layers[1](styles[1]),
+                   # conv layers
+                   self.modulation_layers[2]:  self.modulation_layers[2](styles[2]),
+                   self.modulation_layers[3]:  self.modulation_layers[3](styles[3]),
+                   self.modulation_layers[5]:  self.modulation_layers[5](styles[4]),
+                   self.modulation_layers[6]:  self.modulation_layers[6](styles[5]),
+                   self.modulation_layers[8]:  self.modulation_layers[8](styles[6]),
+                   self.modulation_layers[9]:  self.modulation_layers[9](styles[7]),
+                   self.modulation_layers[11]: self.modulation_layers[11](styles[8]),
+                   self.modulation_layers[12]: self.modulation_layers[12](styles[9]),
+                   self.modulation_layers[14]: self.modulation_layers[14](styles[10]),
+                   self.modulation_layers[15]: self.modulation_layers[15](styles[11]),
+                   self.modulation_layers[17]: self.modulation_layers[17](styles[12]),
+                   self.modulation_layers[18]: self.modulation_layers[18](styles[13]),
+                   self.modulation_layers[20]: self.modulation_layers[20](styles[14]),
+                   self.modulation_layers[21]: self.modulation_layers[21](styles[15]),
+                   self.modulation_layers[23]: self.modulation_layers[23](styles[16]),
+                   self.modulation_layers[24]: self.modulation_layers[24](styles[17]),
+                   # toRGB layers
+                   self.modulation_layers[4]: self.modulation_layers[4](styles[3]),
+                   self.modulation_layers[7]: self.modulation_layers[7](styles[5]),
+                   self.modulation_layers[10]: self.modulation_layers[10](styles[7]),
+                   self.modulation_layers[13]: self.modulation_layers[13](styles[9]),
+                   self.modulation_layers[16]: self.modulation_layers[16](styles[11]),
+                   self.modulation_layers[19]: self.modulation_layers[19](styles[13]),
+                   self.modulation_layers[22]: self.modulation_layers[22](styles[15]),
+                   self.modulation_layers[25]: self.modulation_layers[25](styles[17]),
+        }
 
         return s_codes    
 

styleclip/styleclip_global.py

@@ -120,7 +120,10 @@ def get_direction(neutral_class, target_class, beta, di, clip_model=None):
 
     dt = class_weights[:, 1] - class_weights[:, 0]
     dt = dt / dt.norm()
-    dt = dt.type(type(di))
+
+    dt = dt.float()
+    di = di.float()
+
     relevance = di @ dt
     mask = relevance.abs() > beta
     direction = relevance * mask
@@ -144,7 +147,7 @@ def style_tensor_to_style_dict(style_tensor, refernce_generator):
 def style_dict_to_style_tensor(style_dict, reference_generator):
     style_layers = reference_generator.modulation_layers
 
-    style_tensor = torch.zeros(shape=(1, 9088))
+    style_tensor = torch.zeros(size=(1, 9088))
     for layer in style_dict:
         layer_idx = style_layers.index(layer)
         style_tensor[:, FFHQ_CODE_INDICES[layer_idx][0]:FFHQ_CODE_INDICES[layer_idx][1]] = style_dict[layer]