import gradio as gr
import numpy as np
from transformers import BeitFeatureExtractor, BeitForSemanticSegmentation, DPTFeatureExtractor, DPTForDepthEstimation
from PIL import Image, ImageFilter
import requests
import torch
import cv2

# Load the segmentation model and feature extractor
segmentation_processor = BeitFeatureExtractor.from_pretrained("microsoft/beit-base-finetuned-ade-640-640")
segmentation_model = BeitForSemanticSegmentation.from_pretrained("microsoft/beit-base-finetuned-ade-640-640")

# Load the depth estimation model and feature extractor
depth_feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-large")
depth_model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large")

def apply_gaussian_blur(image):
    # Resize and preprocess the image
    image = image.resize((512, 512)).convert("RGB")
    inputs = segmentation_processor(image, return_tensors="pt")

    # Perform segmentation to get the 'person' mask
    with torch.no_grad():
        outputs = segmentation_model(**inputs)
        logits = outputs.logits
    segmentation = torch.argmax(logits, dim=1)[0]
    
    # Create a binary mask for the 'person' class (index 12)
    person_index = 12
    binary_mask = (segmentation == person_index).numpy().astype(np.uint8) * 255

    # Apply Gaussian blur to the entire image
    image_np = np.array(image)
    blurred_image = cv2.GaussianBlur(image_np, (0, 0), sigmaX=15, sigmaY=15)

    # Normalize the mask to range between 0 and 1
    normalized_mask = binary_mask / 255.0
    normalized_mask = np.expand_dims(normalized_mask, axis=-1)

    # Create the composite image
    final_image = (image_np * normalized_mask + blurred_image * (1 - normalized_mask)).astype(np.uint8)

    return Image.fromarray(final_image)

def apply_lens_blur(image):
    # Resize and preprocess the image
    image = image.resize((512, 512)).convert("RGB")
    depth_inputs = depth_feature_extractor(images=image, return_tensors="pt")

    # Perform depth estimation
    with torch.no_grad():
        depth_outputs = depth_model(**depth_inputs)
        predicted_depth = depth_outputs.predicted_depth[0].cpu().numpy()

    # Normalize and invert the depth map
    min_depth = predicted_depth.min()
    max_depth = predicted_depth.max()
    normalized_depth = (predicted_depth - min_depth) / (max_depth - min_depth)
    inverted_depth = 1 - normalized_depth

    # Resize the depth map to match the original image size
    depth_weight_resized = Image.fromarray((inverted_depth * 255).astype(np.uint8)).resize((512, 512))
    depth_weight_resized = np.array(depth_weight_resized) / 255.0
    depth_weight_resized = depth_weight_resized[:, :, np.newaxis]

    # Apply maximum Gaussian blur to the original image
    blurred_image = image.filter(ImageFilter.GaussianBlur(radius=15))

    # Convert images to numpy arrays
    original_np = np.array(image).astype(np.float32)
    blurred_np = np.array(blurred_image).astype(np.float32)

    # Blend the images based on the resized depth map
    output_np = (1 - depth_weight_resized) * original_np + depth_weight_resized * blurred_np

    # Convert back to uint8
    output_np = np.clip(output_np, 0, 255).astype(np.uint8)

    return Image.fromarray(output_np)

# Define the Gradio interface
interface = gr.Interface(
    fn={"Gaussian Blur": apply_gaussian_blur, "Lens Blur": apply_lens_blur},
    inputs=gr.inputs.Image(type="pil"),
    outputs=gr.outputs.Image(type="pil"),
    title="Blur Effects with Hugging Face",
    description="Apply Gaussian Blur or Lens Blur to images using semantic segmentation or depth estimation."
)

# Launch the Gradio interface
interface.launch()