Making changes in random seed value generation

app.py

@@ -50,8 +50,10 @@ def init_transformers(device):
 # Add after init_transformers and before generate_images
 def image_loss(images, loss_type, device, elastic_transformer):
     if loss_type == 'blue':
-        error = torch.abs(images[:,2] - 0.9).mean()
-        return error.to(device)
+        # Reduced target blue value from 0.9 to 0.6 for more subtle effect
+        error = torch.abs(images[:,2] - 0.6).mean()
+        # Apply a lower scale specifically for blue loss
+        return (error * 0.3).to(device)  # Reduced scaling factor
     elif loss_type == 'elastic':
         transformed_imgs = elastic_transformer(images)
         error = torch.abs(transformed_imgs - images).mean()
@@ -69,8 +71,8 @@ def image_loss(images, loss_type, device, elastic_transformer):
 
 # Update configuration for faster generation
 height, width = 384, 384  # Reduced from 512x512 to 384x384
-guidance_scale = 7.5
-num_inference_steps = 30
+guidance_scale = 8  # Increased from 7.5 to 8 for better prompt adherence
+num_inference_steps = 45  # Using 45 steps for better quality
 loss_scale = 150
 
 def generate_images(prompt, concept):
@@ -88,6 +90,10 @@ def generate_images(prompt, concept):
     loss_functions = ['none', 'blue', 'elastic', 'symmetry', 'saturation']
     progress = gr.Progress()
     
+    # Generate one random seed for all loss types
+    random_seed = torch.randint(1, 10000, (1,)).item()  # Random seed between 1 and 9999
+    print(f"\nUsing random seed {random_seed} for all images")
+    
     for idx, loss_type in enumerate(loss_functions):
         try:
             print(f"\n[{loss_type.upper()}] Starting image generation...")
@@ -139,8 +145,9 @@ def generate_images(prompt, concept):
             
             text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
             
-            # Generate initial latents with correct dtype
-            generator = torch.manual_seed(idx * 1000)
+            # Generate initial latents with random seed
+            # Use the same seed for all loss types
+            generator = torch.manual_seed(random_seed)
             latents = torch.randn(
                 (1, pipe.unet.config.in_channels, height // 8, width // 8),
                 generator=generator,