scscs

app.py

@@ -1,53 +1,67 @@
-import gradio as gr
+import os
 import torch
+import gradio as gr
 from diffusers import StableDiffusionImg2ImgPipeline
 from PIL import Image
-import os
-# Check and print device status
-os.environ["CUDA_VISIBLE_DEVICES"] = "0"  # Ensure CUDA is used
+
+# Force CUDA usage
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+torch.backends.cudnn.benchmark = True
+torch.backends.cuda.matmul.allow_tf32 = True
+
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 print(f"Using device: {device}")
-print(f"CUDA available: {torch.cuda.is_available()}")
 
-# Ensure torch is installed correctly
+# Ensure torch is properly installed
 try:
     torch.zeros(1).to(device)
     print("Torch initialized successfully on", device)
 except Exception as e:
     print("Torch initialization error:", e)
 
-# Load the Stable Diffusion model with optimizations
+# Load the optimized Stable Diffusion model
 model_id = "nitrosocke/Ghibli-Diffusion"
+
 pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
-    model_id, 
-    torch_dtype=torch.float32, 
+    model_id,
+    torch_dtype=torch.float16,
+    use_safetensors=True,
+    low_cpu_mem_usage=True,
     safety_checker=None
 ).to(device)
 
-if device == "cuda":
-    pipe.to(device)
-    pipe.enable_model_cpu_offload()  # Efficient VRAM usage
-    pipe.enable_xformers_memory_efficient_attention()  # Optimized attention for speed
+pipe.enable_xformers_memory_efficient_attention()
+pipe.enable_model_cpu_offload()
+pipe.enable_vae_slicing()
+pipe.enable_attention_slicing()
 
-def transform_image(input_image: Image.Image) -> Image.Image:
+# Enhanced prompt for Studio Ghibli-style transformation
+prompt = (
+    "Beautiful Studio Ghibli anime-style portrait, breathtaking landscape background, "
+    "soft pastel colors, hand-painted texture, cinematic lighting, dreamy atmosphere, "
+    "vibrant and rich details, Miyazaki aesthetic, magical realism, watercolor effect, "
+    "warm sunlight, stunning composition, high detail, fantasy world."
+)
+
+def transform_image(input_image):
     input_image = input_image.resize((512, 512))
-    prompt = "ghibli style, cinematic lighting, hand-painted, anime aesthetics"
 
     output = pipe(
         prompt=prompt,
         image=input_image,
-        strength=0.65,  # Reduce strength to avoid excessive details
-        guidance_scale=5.0,  # Lowered for faster inference
-        num_inference_steps=25,  # Reduced steps for speed
+        strength=0.65,
+        guidance_scale=4.5,
+        num_inference_steps=20,
     )
+
     return output.images[0]
 
-# Gradio Interface
+# Gradio UI
 demo = gr.Interface(
     fn=transform_image,
     inputs=gr.Image(type="pil", label="Upload your portrait/photo"),
     outputs=gr.Image(type="pil", label="Studio Ghibli Style Output"),
-    title="Studio Ghibli Style Converter",
+    title="Studio Ghibli AI Converter",
     description="Upload a portrait or photo to transform it into a Studio Ghibli-style image.",
 )