app.py

import cv2
import numpy as np
from PIL import Image, ImageFilter
import torch
import gradio as gr
from torchvision import transforms
from transformers import (
    AutoModelForImageSegmentation,
    DepthProImageProcessorFast,
    DepthProForDepthEstimation,
)

# Set device
device = "cuda" if torch.cuda.is_available() else "cpu"

# -----------------------------
# Load Segmentation Model (RMBG-2.0 by briaai)
# -----------------------------
seg_model = AutoModelForImageSegmentation.from_pretrained(
    "briaai/RMBG-2.0", trust_remote_code=True
)
torch.set_float32_matmul_precision(["high", "highest"][0])
seg_model.to(device)
seg_model.eval()

# Define segmentation image size and transform
seg_image_size = (1024, 1024)
seg_transform = transforms.Compose([
    transforms.Resize(seg_image_size),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

# -----------------------------
# Load Depth Estimation Model (DepthPro by Apple)
# -----------------------------
depth_processor = DepthProImageProcessorFast.from_pretrained("apple/DepthPro-hf")
depth_model = DepthProForDepthEstimation.from_pretrained("apple/DepthPro-hf")
depth_model.to(device)
depth_model.eval()

# -----------------------------
# Define the Segmentation-Based Blur Effect
# -----------------------------
def segmentation_blur_effect(input_image: Image.Image):
    """
    Creates a segmentation mask using RMBG-2.0 and applies a Gaussian blur (sigma=15)
    to the background while keeping the foreground sharp.
    """
    # Resize input image for segmentation processing
    imageResized = input_image.resize(seg_image_size)
    input_tensor = seg_transform(imageResized).unsqueeze(0).to(device)
    
    with torch.no_grad():
        preds = seg_model(input_tensor)[-1].sigmoid().cpu()
    pred = preds[0].squeeze()
    
    # Convert predicted mask to a PIL image and ensure it matches imageResized's size
    pred_pil = transforms.ToPILImage()(pred)
    mask = pred_pil.resize(imageResized.size)
    
    # Convert mask to grayscale and threshold to create a binary mask
    mask_np = np.array(mask.convert("L"))
    _, maskBinary = cv2.threshold(mask_np, 127, 255, cv2.THRESH_BINARY)
    
    # Convert the resized image to an OpenCV BGR array
    img = cv2.cvtColor(np.array(imageResized), cv2.COLOR_RGB2BGR)
    # Apply Gaussian blur (sigmaX=15, sigmaY=15)
    blurredBg = cv2.GaussianBlur(np.array(imageResized), (0, 0), sigmaX=15, sigmaY=15)
    
    # Create the inverse mask and convert it to 3 channels
    maskInv = cv2.bitwise_not(maskBinary)
    maskInv3 = cv2.cvtColor(maskInv, cv2.COLOR_GRAY2BGR)
    
    # Extract the foreground and background using the mask
    foreground = cv2.bitwise_and(img, cv2.bitwise_not(maskInv3))
    background = cv2.bitwise_and(blurredBg, maskInv3)
    
    # Combine foreground and background; convert back to RGB for display
    finalImg = cv2.add(cv2.cvtColor(foreground, cv2.COLOR_BGR2RGB), background)
    finalImg_pil = Image.fromarray(finalImg)
    
    return finalImg_pil, mask

# -----------------------------
# Define the Depth-Based Lens Blur Effect with Slider-Controlled Thresholds
# -----------------------------
def lens_blur_effect(input_image: Image.Image, fg_threshold: float = 85, mg_threshold: float = 170):
    """
    Uses DepthPro to estimate a depth map and applies a dynamic lens blur effect
    by blending three versions of the image with increasing blur levels.
    
    Parameters:
      input_image: The original PIL image.
      fg_threshold: Foreground threshold (0-255). Pixels with depth below this are considered foreground.
      mg_threshold: Middleground threshold (0-255). Pixels with depth between fg_threshold and mg_threshold are middleground.
    
    Returns:
      depthImg: The computed depth map (PIL Image).
      lensBlurImage: The final lens-blurred image (PIL Image).
      mask_fg_img: Foreground depth mask.
      mask_mg_img: Middleground depth mask.
      mask_bg_img: Background depth mask.
    """
    # Process the image with the depth estimation model
    inputs = depth_processor(images=input_image, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = depth_model(**inputs)
    post_processed_output = depth_processor.post_process_depth_estimation(
        outputs, target_sizes=[(input_image.height, input_image.width)]
    )
    depth = post_processed_output[0]["predicted_depth"]
    
    # Normalize depth to [0, 255]
    depth = (depth - depth.min()) / (depth.max() - depth.min())
    depth = depth * 255.
    depth = depth.detach().cpu().numpy()
    depth_map = depth.astype(np.uint8)
    depthImg = Image.fromarray(depth_map)
    
    # Convert input image to OpenCV BGR format
    img = cv2.cvtColor(np.array(input_image), cv2.COLOR_RGB2BGR)
    
    # Precompute blurred versions for different depth regions
    img_foreground = img.copy()  # No blur for foreground
    img_middleground = cv2.GaussianBlur(img, (0, 0), sigmaX=7, sigmaY=7)
    img_background = cv2.GaussianBlur(img, (0, 0), sigmaX=15, sigmaY=15)

    print(depth_map)
    depth_map /= depth_map.max()
    
    # Use slider values as thresholds
    threshold1 = fg_threshold   # e.g., default 85
    threshold2 = mg_threshold   # e.g., default 170
    
    # Create masks for foreground, middleground, and background based on depth
    mask_fg = (depth_map < threshold1).astype(np.float32)
    mask_mg = ((depth_map >= threshold1) & (depth_map < threshold2)).astype(np.float32)
    mask_bg = (depth_map >= threshold2).astype(np.float32)
    
    # Expand masks to 3 channels
    mask_fg_3 = np.stack([mask_fg]*3, axis=-1)
    mask_mg_3 = np.stack([mask_mg]*3, axis=-1)
    mask_bg_3 = np.stack([mask_bg]*3, axis=-1)
    
    # Blend the images using the masks
    final_img = (img_foreground * mask_fg_3 +
                 img_middleground * mask_mg_3 +
                 img_background * mask_bg_3).astype(np.uint8)
    
    final_img_rgb = cv2.cvtColor(final_img, cv2.COLOR_BGR2RGB)
    lensBlurImage = Image.fromarray(final_img_rgb)
    
    # Create mask images for display (scaled to 0-255)
    mask_fg_img = Image.fromarray((mask_fg * 255).astype(np.uint8))
    mask_mg_img = Image.fromarray((mask_mg * 255).astype(np.uint8))
    mask_bg_img = Image.fromarray((mask_bg * 255).astype(np.uint8))
    
    return depthImg, lensBlurImage, mask_fg_img, mask_mg_img, mask_bg_img

# -----------------------------
# Gradio App: Process Image and Display Multiple Effects
# -----------------------------
def process_image(input_image: Image.Image, fg_threshold: float, mg_threshold: float):
    """
    Processes the uploaded image to generate:
      1. Segmentation-based Gaussian blur effect.
      2. Segmentation mask.
      3. Depth map.
      4. Depth-based lens blur effect.
      5. Depth masks for foreground, middleground, and background.
    
    The depth thresholds for foreground and middleground regions are adjustable via sliders.
    """
    seg_blur, seg_mask = segmentation_blur_effect(input_image)
    depth_map_img, lens_blur_img, mask_fg_img, mask_mg_img, mask_bg_img = lens_blur_effect(
        input_image, fg_threshold, mg_threshold
    )
    
    return (
        seg_blur,
        seg_mask,
        depth_map_img,
        lens_blur_img,
        mask_fg_img,
        mask_mg_img,
        mask_bg_img
    )

title = "Blur Effects: Gaussian & Depth-Based Lens Blur with Adjustable Depth Thresholds"
description = (
    "Upload an image to apply two distinct effects:\n\n"
    "1. A segmentation-based Gaussian blur that blurs the background (using RMBG-2.0).\n"
    "2. A depth-based lens blur effect that simulates realistic lens blur based on depth (using DepthPro).\n\n"
    "Use the sliders to adjust the foreground and middleground depth thresholds."
)

demo = gr.Interface(
    fn=process_image,
    inputs=[
        gr.Image(type="pil", label="Input Image"),
        gr.Slider(minimum=0, maximum=1, step=0.01, value=0.33, label="Foreground Depth Threshold"),
        gr.Slider(minimum=0, maximum=1, step=0.01, value=0.66, label="Middleground Depth Threshold")
    ],
    outputs=[
        gr.Image(type="pil", label="Segmentation-Based Blur"),
        gr.Image(type="pil", label="Segmentation Mask"),
        gr.Image(type="pil", label="Depth Map"),
        gr.Image(type="pil", label="Depth-Based Lens Blur"),
        gr.Image(type="pil", label="Foreground Depth Mask"),
        gr.Image(type="pil", label="Middleground Depth Mask"),
        gr.Image(type="pil", label="Background Depth Mask")
    ],
    title=title,
    description=description,
    allow_flagging="never"
)

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