app.py

import io
import numpy as np
import torch
from PIL import Image, ImageFilter
from torchvision import transforms
import gradio as gr
from transformers import AutoModelForImageSegmentation, pipeline

# ----------------------------
# Global Setup and Model Loading
# ----------------------------

# Set device (GPU if available, else CPU)
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load the segmentation model (RMBG-2.0)
segmentation_model = AutoModelForImageSegmentation.from_pretrained(
    'briaai/RMBG-2.0',
    trust_remote_code=True
)
segmentation_model.to(device)
segmentation_model.eval()

# Define the image transformation for segmentation (resize to 512x512)
image_size = (512, 512)
segmentation_transform = transforms.Compose([
    transforms.Resize(image_size),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

# Load the depth estimation pipeline (Depth-Anything)
depth_pipeline = pipeline("depth-estimation", model="depth-anything/Depth-Anything-V2-Small-hf")


# ----------------------------
# Processing Functions
# ----------------------------

def segment_and_blur_background(input_image: Image.Image, blur_radius: int = 15, threshold: float = 0.5) -> Image.Image:
    """
    Applies segmentation using the RMBG-2.0 model and then uses the segmentation mask
    to composite a Gaussian-blurred background with a sharp foreground.
    """
    # Ensure the image is in RGB and get its original dimensions
    image = input_image.convert("RGB")
    orig_width, orig_height = image.size

    # Preprocess image for segmentation
    input_tensor = segmentation_transform(image).unsqueeze(0).to(device)
    
    # Run inference on the segmentation model
    with torch.no_grad():
        preds = segmentation_model(input_tensor)[-1].sigmoid().cpu()
    pred = preds[0].squeeze()
    
    # Create a binary mask using the threshold
    binary_mask = (pred > threshold).float()
    mask_pil = transforms.ToPILImage()(binary_mask).convert("L")
    # Convert grayscale mask to pure binary (0 or 255)
    mask_pil = mask_pil.point(lambda p: 255 if p > 128 else 0)
    # Resize mask back to the original image dimensions
    mask_pil = mask_pil.resize((orig_width, orig_height), resample=Image.BILINEAR)
    
    # Apply Gaussian blur to the entire image for background
    blurred_image = image.filter(ImageFilter.GaussianBlur(blur_radius))
    # Composite the original image (foreground) with the blurred image (background) using the mask
    final_image = Image.composite(image, blurred_image, mask_pil)
    return final_image


def depth_based_lens_blur(input_image: Image.Image, max_blur: float = 2, num_bands: int = 40, invert_depth: bool = False) -> Image.Image:
    """
    Applies a depth-based blur effect using a depth map produced by Depth-Anything.
    The effect simulates a lens blur by applying different blur strengths in depth bands.
    """
    # Resize the input image to 512x512 for the depth estimation model
    image_resized = input_image.resize((512, 512))
    
    # Run depth estimation to obtain the depth map (as a PIL image)
    results = depth_pipeline(image_resized)
    depth_map_image = results['depth']
    
    # Convert the depth map to a NumPy array and normalize to [0, 1]
    depth_array = np.array(depth_map_image, dtype=np.float32)
    d_min, d_max = depth_array.min(), depth_array.max()
    depth_norm = (depth_array - d_min) / (d_max - d_min + 1e-8)
    if invert_depth:
        depth_norm = 1.0 - depth_norm

    # Convert the resized image to RGBA for compositing
    orig_rgba = image_resized.convert("RGBA")
    final_image = orig_rgba.copy()
    
    # Divide the normalized depth range into bands
    band_edges = np.linspace(0, 1, num_bands + 1)
    for i in range(num_bands):
        band_min = band_edges[i]
        band_max = band_edges[i + 1]
        # Use the midpoint of the band to determine the blur strength.
        mid = (band_min + band_max) / 2.0
        blur_radius_band = (1 - mid) * max_blur
        
        # Create a blurred version of the image for this band.
        blurred_version = orig_rgba.filter(ImageFilter.GaussianBlur(blur_radius_band))
        
        # Create a mask for pixels whose normalized depth falls within this band.
        band_mask = ((depth_norm >= band_min) & (depth_norm < band_max)).astype(np.uint8) * 255
        band_mask_pil = Image.fromarray(band_mask, mode="L")
        
        # Composite the blurred version with the current final image using the band mask.
        final_image = Image.composite(blurred_version, final_image, band_mask_pil)
    
    # Return the final composited image as RGB.
    return final_image.convert("RGB")


def process_image(input_image: Image.Image, effect: str) -> Image.Image:
    """
    Dispatch function to apply the selected effect:
    - "Gaussian Blur Background": uses segmentation and Gaussian blur.
    - "Depth-based Lens Blur": applies depth-based blur using the estimated depth map.
    """
    if effect == "Gaussian Blur Background":
        return segment_and_blur_background(input_image)
    elif effect == "Depth-based Lens Blur":
        return depth_based_lens_blur(input_image)
    else:
        return input_image


# ----------------------------
# Gradio Interface
# ----------------------------

iface = gr.Interface(
    fn=process_image,
    inputs=[
        gr.Image(type="pil", label="Input Image"),
        gr.Radio(choices=["Gaussian Blur Background", "Depth-based Lens Blur"], label="Select Effect")
    ],
    outputs=gr.Image(type="pil", label="Output Image"),
    title="Blur Effects Demo",
    description=(
        "Upload an image and choose an effect: "
        "apply segmentation + Gaussian blurred background, or a depth-based lens blur effect."
    )
)

if __name__ == "__main__":
    iface.launch()