Update app.py

app.py

@@ -4,17 +4,33 @@ from PIL import Image
 from diffusers import AutoPipelineForText2Image, DDIMScheduler
 from transformers import CLIPVisionModelWithProjection
 import numpy as np
-import spaces  # Ensure this is available in your environment
 
-# Initialize a zero tensor for demonstration purposes
-zero = torch.Tensor([0]).cuda()
-print(zero.device)  # Should output 'cuda:0' if a GPU is available
+# Initialize the pipeline with GPU support
+pipeline = AutoPipelineForText2Image.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16,
+    device="cuda",  # Use GPU device if available
+)
+
+# Configure the scheduler for the pipeline
+pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
+
+# Load IP adapter with specified weights and set the scale for each component
+pipeline.load_ip_adapter(
+    "h94/IP-Adapter",
+    subfolder="sdxl_models",
+    weight_name=[
+        "ip-adapter-plus_sdxl_vit-h.safetensors",
+        "ip-adapter-plus-face_sdxl_vit-h.safetensors"
+    ]
+)
+pipeline.set_ip_adapter_scale([0.7, 0.5])
+
+# Ensure the model and its components are moved to GPU
+pipeline.to("cuda")
 
-@spaces.GPU  # Decorate the function to run on GPU
 def transform_image(face_image):
-    print(zero.device)  # Check the device inside the function, should be 'cuda:0'
-    
-    generator = torch.Generator(device="cuda").manual_seed(0)  # Use GPU device if available
+    generator = torch.Generator(device="cuda").manual_seed(0)
 
     # Process the input face image
     if isinstance(face_image, Image.Image):
@@ -25,10 +41,10 @@ def transform_image(face_image):
         raise ValueError("Unsupported image format")
 
     # Load the style image from the local path
-    style_image_path = "/content/soyjak2.jpeg"
+    style_image_path = "InstaSoyjak/soyjak2.jpeg"
     style_image = Image.open(style_image_path)
 
-    # Perform the transformation using the GPU
+    # Perform the transformation using the configured pipeline
     image = pipeline(
         prompt="soyjak",
         ip_adapter_image=[style_image, processed_face_image],
@@ -39,25 +55,14 @@ def transform_image(face_image):
 
     return image
 
-# Load models and configure pipeline with GPU support
-pipeline = AutoPipelineForText2Image.from_pretrained(
-    "stabilityai/stable-diffusion-xl-base-1.0",
-    torch_dtype=torch.float16,  # Consider using torch.float32 for GPU computations
-    device="cuda",  # Use GPU device if available
-).to("cuda")  # Ensure the model is moved to GPU
-
-# Additional pipeline configurations
-pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config).to("cuda")
-pipeline.enable_model_cpu_offload(False)  # Consider not offloading to CPU when using GPU
-
 # Gradio interface setup
 demo = gr.Interface(
     fn=transform_image,
     inputs=gr.Image(label="Upload your face image"),
     outputs=gr.Image(label="Your Soyjak"),
     title="InstaSoyjak - turn anyone into a Soyjak",
-    description="All you need to do is upload an image. Please use responsibly. Please follow me on Twitter if you like this space: https://twitter.com/angrypenguinPNG. Idea from Yacine, please give him a follow: https://twitter.com/yacineMTB.",
+    description="All you need to do is upload an image. Please use responsibly.",
 )
 
-demo.queue(max_size=20)  # Configures the queue with a maximum size of 20
+demo.queue(max_size=20)
 demo.launch()