app.py

import gradio as gr
import torch
import cv2
import numpy as np
from torchvision import transforms
from PIL import Image
from scipy.interpolate import Rbf

# Load MiDaS depth estimation model
midas_model = torch.hub.load("intel-isl/MiDaS", "DPT_Hybrid")
midas_model.eval()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
midas_model.to(device)
midas_transform = torch.hub.load("intel-isl/MiDaS", "transforms").default_transform

def estimate_depth(image):
    image = image.convert("RGB")
    image_np = np.array(image) / 255.0  # Convert PIL image to NumPy and normalize
    image_tensor = torch.tensor(image_np, dtype=torch.float32).permute(2, 0, 1).unsqueeze(0).to(device)
    
    with torch.no_grad():
        depth = midas_model(image_tensor).squeeze().cpu().numpy()
    
    depth = cv2.resize(depth, (image.size[0], image.size[1]))
    depth = (depth - depth.min()) / (depth.max() - depth.min()) * 255
    return depth.astype(np.uint8)

def compute_optical_flow(depth):
    depth_blurred = cv2.GaussianBlur(depth, (5, 5), 0)
    flow = cv2.calcOpticalFlowFarneback(depth_blurred, depth, None, 0.5, 3, 15, 3, 5, 1.2, 0)
    displacement_x = cv2.normalize(flow[..., 0], None, -5, 5, cv2.NORM_MINMAX)
    displacement_y = cv2.normalize(flow[..., 1], None, -5, 5, cv2.NORM_MINMAX)
    return displacement_x, displacement_y

def apply_tps_interpolation(design, depth):
    h, w = depth.shape
    grid_x, grid_y = np.meshgrid(np.arange(w), np.arange(h))
    edges = cv2.Canny(depth.astype(np.uint8), 50, 150)
    points = np.column_stack(np.where(edges > 0))
    tps_x = Rbf(points[:, 1], points[:, 0], grid_x[points[:, 0], points[:, 1]], function="thin_plate")
    tps_y = Rbf(points[:, 1], points[:, 0], grid_y[points[:, 0], points[:, 1]], function="thin_plate")
    map_x = tps_x(grid_x, grid_y).astype(np.float32)
    map_y = tps_y(grid_x, grid_y).astype(np.float32)
    return cv2.remap(design, map_x, map_y, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_REFLECT)

def compute_adaptive_alpha(depth):
    grad_x = cv2.Sobel(depth, cv2.CV_32F, 1, 0, ksize=3)
    grad_y = cv2.Sobel(depth, cv2.CV_32F, 0, 1, ksize=3)
    grad_magnitude = np.sqrt(grad_x**2 + grad_y**2)
    alpha = cv2.normalize(grad_magnitude, None, 0, 1, cv2.NORM_MINMAX)
    return alpha

def blend_design(cloth_img, design_img):
    cloth_img = cloth_img.convert("RGB")
    design_img = design_img.convert("RGBA")
    cloth_np = np.array(cloth_img)
    design_np = np.array(design_img)
    h, w, _ = cloth_np.shape
    dh, dw, _ = design_np.shape
    scale_factor = min(w / dw, h / dh) * 0.4  
    new_w, new_h = int(dw * scale_factor), int(dh * scale_factor)
    design_np = cv2.resize(design_np, (new_w, new_h), interpolation=cv2.INTER_AREA)
    alpha_channel = design_np[:, :, 3] / 255.0
    design_np = design_np[:, :, :3]
    x_offset = (w - new_w) // 2
    y_offset = int(h * 0.35)
    design_canvas = np.zeros_like(cloth_np)
    design_canvas[y_offset:y_offset+new_h, x_offset:x_offset+new_w] = design_np
    depth_map = estimate_depth(cloth_img)
    warped_design = apply_tps_interpolation(design_canvas, depth_map)
    adaptive_alpha = compute_adaptive_alpha(depth_map)
    cloth_np = (cloth_np * (1 - adaptive_alpha) + warped_design * adaptive_alpha).astype(np.uint8)
    return Image.fromarray(cloth_np)

def main(cloth, design):
    global midas_model
    if midas_model is None:
        midas_model = torch.hub.load("intel-isl/MiDaS", "MiDaS_small").to(device).eval()
    return blend_design(cloth, design)

iface = gr.Interface(
    fn=main,
    inputs=[gr.Image(type="pil"), gr.Image(type="pil")],
    outputs=gr.Image(type="pil"),
    title="AI Cloth Design Warping",
    description="Upload a clothing image and a design to blend it naturally, ensuring it stays centered and follows fabric folds."
)

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