Update app.py

app.py

@@ -6,22 +6,21 @@ from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
 from functools import lru_cache
 from PIL import Image
 
-from torchvision import transforms
 from transformers import CLIPImageProcessor  # Updated per deprecation warning
-import os
+
 
 
 @lru_cache(maxsize=1)
 def load_pipeline():
-    # Determine device and set torch_dtype accordingly
+    # Determine device and appropriate torch_dtype
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     torch_dtype = torch.bfloat16 if device.type == "cuda" else torch.float32
 
-    # Load base model with the appropriate dtype
     base_model = "black-forest-labs/FLUX.1-dev"
     pipe = DiffusionPipeline.from_pretrained(
         base_model,
-        torch_dtype=torch_dtype
+        torch_dtype=torch_dtype,
+        low_cpu_mem_usage=True  # Reduce memory usage during load
     )
     
     # Load LoRA weights
@@ -34,12 +33,17 @@ def load_pipeline():
     )
     image_processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch32")
     
-    # Enable GPU-only optimizations if a GPU is available
+    # Enable GPU optimizations if on GPU; else, try sequential offloading on CPU
     if device.type == "cuda":
         try:
             pipe.enable_xformers_memory_efficient_attention()
         except Exception as e:
             print("Warning: Could not enable xformers memory efficient attention:", e)
+    else:
+        try:
+            pipe.enable_sequential_cpu_offload()
+        except Exception as e:
+            print("Warning: Could not enable sequential CPU offload:", e)
     
     pipe = pipe.to(device)
     return pipe, safety_checker, image_processor
@@ -49,8 +53,8 @@ pipe, safety_checker, image_processor = load_pipeline()
 def generate_image(
     prompt,
     seed=42,
-    width=1024,
-    height=1024,
+    width=512,   # Lowered default resolution
+    height=512,  # Lowered default resolution
     guidance_scale=6,
     steps=28,
     progress=gr.Progress()
@@ -60,11 +64,10 @@ def generate_image(
         device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         generator = torch.Generator(device=device).manual_seed(seed)
         
-        # Ensure the trigger word is present
+        # Auto-add the trigger word if not already present
         if "super realism" not in prompt.lower():
             prompt = f"Super Realism, {prompt}"
         
-        # Define a callback to update progress
         def update_progress(step, timestep, latents):
             progress((step + 1) / steps, desc="Generating image...")
         
@@ -81,13 +84,12 @@ def generate_image(
             image = result.images[0]
         
         progress(1, desc="Safety checking...")
-        # Preprocess the image for safety checking
         safety_input = image_processor(image, return_tensors="pt")
         np_image = np.array(image)
         
-        # Run the safety checker
+        # Run safety checker; it returns a tuple where the second element is nsfw flags
         _, nsfw_detected = safety_checker(
-            images=[np_image], 
+            images=[np_image],
             clip_input=safety_input.pixel_values
         )
         
@@ -106,8 +108,9 @@ with gr.Blocks() as app:
         with gr.Column():
             prompt_input = gr.Textbox(label="Prompt", value="A portrait of a person")
             seed_input = gr.Slider(0, 1000, value=42, label="Seed")
-            width_input = gr.Slider(512, 2048, value=1024, label="Width")
-            height_input = gr.Slider(512, 2048, value=1024, label="Height")
+            # Lower the resolution slider range for less memory-intensive generation
+            width_input = gr.Slider(256, 1024, value=512, label="Width")
+            height_input = gr.Slider(256, 1024, value=512, label="Height")
             guidance_input = gr.Slider(1, 20, value=6, label="Guidance Scale")
             steps_input = gr.Slider(10, 100, value=28, label="Steps")
             submit = gr.Button("Generate")
@@ -122,5 +125,5 @@ with gr.Blocks() as app:
         outputs=[output_image, status]
     )
     
-    # Queue without GPU-specific arguments
+    # Use queue without GPU-specific parameters
     app.queue(max_size=3).launch()