app.py

import gradio as gr
import torch
import cv2
import numpy as np
from torchvision import transforms
from PIL import Image

# Load MiDaS depth estimation model from torch.hub
midas_model = torch.hub.load("intel-isl/MiDaS", "MiDaS_small")
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):
    """Estimate depth map using MiDaS v3."""
    image = image.convert("RGB")  # Ensure it's in RGB format

    # Convert PIL image to a NumPy array and normalize it
    img_np = np.array(image, dtype=np.float32) / 255.0  # Normalize to [0, 1]

    # Convert NumPy array to a Torch tensor
    img_tensor = torch.tensor(img_np).permute(2, 0, 1).unsqueeze(0).to(device)

    # Ensure tensor shape is [1, 3, H, W]
    if img_tensor.dim() == 5:  # If an extra batch dimension is present
        img_tensor = img_tensor.squeeze(1)

    with torch.no_grad():
        depth = midas_model(img_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 apply_tps_warping(design, depth):
    """Apply Thin Plate Spline (TPS) warping based on depth."""
    h, w = depth.shape
    grid_x, grid_y = np.meshgrid(np.arange(w), np.arange(h))
    displacement_x = cv2.Sobel(depth, cv2.CV_32F, 1, 0, ksize=5)
    displacement_y = cv2.Sobel(depth, cv2.CV_32F, 0, 1, ksize=5)
    displacement_x = cv2.normalize(displacement_x, None, -5, 5, cv2.NORM_MINMAX)
    displacement_y = cv2.normalize(displacement_y, None, -5, 5, cv2.NORM_MINMAX)
    
    map_x = np.clip(grid_x + displacement_x, 0, w - 1).astype(np.float32)
    map_y = np.clip(grid_y + displacement_y, 0, h - 1).astype(np.float32)
    
    warped_design = cv2.remap(design, map_x, map_y, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_REFLECT)
    return warped_design

def blend_design(cloth_img, design_img):
    """Blend design onto clothing naturally with fold adaptation using TPS warping."""
    cloth_img = cloth_img.convert("RGB")
    design_img = design_img.convert("RGBA")
    cloth_np = np.array(cloth_img)
    design_np = np.array(design_img)
    
    # Resize design
    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)
    
    # Extract alpha channel
    alpha_channel = design_np[:, :, 3] / 255.0
    design_np = design_np[:, :, :3]
    
    # Create placement area
    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
    
    # Estimate depth and apply TPS warping
    depth_map = estimate_depth(cloth_img)
    warped_design = apply_tps_warping(design_canvas, depth_map)
    
    # Ensure alpha is applied correctly
    mask = np.zeros_like(cloth_np, dtype=np.float32)
    mask[y_offset:y_offset+new_h, x_offset:x_offset+new_w] = np.expand_dims(alpha_channel, axis=-1)
    cloth_np = (cloth_np * (1 - mask) + warped_design * mask).astype(np.uint8)
    
    return Image.fromarray(cloth_np)

def main(cloth, design):
    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=True)