generate.py

import os
import argparse
import cv2
import numpy as np
import torch
from utils.imgops import esrgan_launcher_split_merge, crop_seamless

# Define model paths
NORMAL_MAP_MODEL = 'utils/models/1x_NormalMapGenerator-CX-Lite_200000_G.pth'
OTHER_MAP_MODEL = 'utils/models/1x_FrankenMapGenerator-CX-Lite_215000_G.pth'

def process(img, model, device=None):
    """
    Process an image through the model to generate material maps.
    
    Args:
        img: Input image (numpy array)
        model: PyTorch model to process the image
        device: Torch device to use (defaults to CPU if None)
    
    Returns:
        Processed output image (numpy array)
    """
    if device is None:
        device = torch.device('cpu')  # Default to CPU if no device provided
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = img * 1.0 / 255
    img = torch.from_numpy(np.transpose(img, (2, 0, 1))).float()
    img_LR = img.unsqueeze(0)
    img_LR = img_LR.to(device)
    with torch.no_grad():
        output = model(img_LR).data.squeeze().float().cpu().clamp_(0, 1).numpy()
    output = np.transpose(output, (1, 2, 0))
    output = (output * 255.0).round()
    return output

def load_model(model_path, device=None):
    """
    Load a pre-trained model from a file.
    
    Args:
        model_path: Path to the model file (.pth)
        device: Torch device to use (defaults to CPU if None)
    
    Returns:
        Loaded PyTorch model
    """
    if device is None:
        device = torch.device('cpu')  # Default to CPU if no device provided
    from utils.architecture.architecture import RRDB  # Corrected import
    model = RRDB(nb=15, gc=32)
    model.load_state_dict(torch.load(model_path), strict=True)
    model.eval()
    return model.to(device)

def main():
    # Parse command-line arguments
    parser = argparse.ArgumentParser(description="Generate material maps from diffuse textures.")
    parser.add_argument('--input', type=str, default='input', help='Input directory or image path')
    parser.add_argument('--output', type=str, default='output', help='Output directory')
    parser.add_argument('--tile_size', type=int, default=512, help='Tile size for processing large images')
    parser.add_argument('--seamless', action='store_true', help='Enable seamless tiling with wrap')
    parser.add_argument('--mirror', action='store_true', help='Enable seamless tiling with mirror')
    parser.add_argument('--replicate', action='store_true', help='Enable seamless tiling with replicate')
    parser.add_argument('--ishiiruka', action='store_true', help='Output in Ishiiruka format')
    parser.add_argument('--ishiiruka_texture_encoder', action='store_true', help='Output in Ishiiruka texture encoder format')
    parser.add_argument('--cpu', action='store_true', help='Force CPU usage')
    args = parser.parse_args()

    # Set device based on args.cpu
    device = torch.device('cpu' if args.cpu else 'cuda')

    # Load models
    models = [
        load_model(NORMAL_MAP_MODEL, device),
        load_model(OTHER_MAP_MODEL, device),
    ]

    # Ensure output directory exists
    if not os.path.exists(args.output):
        os.makedirs(args.output)

    # Process each image in the input directory
    if os.path.isdir(args.input):
        for filename in os.listdir(args.input):
            if filename.lower().endswith(('.png', '.jpg', '.jpeg', '.bmp', '.tga')):
                img_path = os.path.join(args.input, filename)
                base = os.path.splitext(filename)[0]
                process_image(img_path, base, args, models, device)
    else:
        base = os.path.splitext(os.path.basename(args.input))[0]
        process_image(args.input, base, args, models, device)

def process_image(img_path, base, args, models, device):
    """Helper function to process a single image."""
    img = cv2.imread(img_path, cv2.IMREAD_COLOR)

    if args.seamless:
        img = cv2.copyMakeBorder(img, 16, 16, 16, 16, cv2.BORDER_WRAP)
    elif args.mirror:
        img = cv2.copyMakeBorder(img, 16, 16, 16, 16, cv2.BORDER_REFLECT_101)
    elif args.replicate:
        img = cv2.copyMakeBorder(img, 16, 16, 16, 16, cv2.BORDER_REPLICATE)

    img_height, img_width = img.shape[:2]
    do_split = img_height > args.tile_size or img_width > args.tile_size

    if do_split:
        rlts = esrgan_launcher_split_merge(img, lambda x, m: process(x, m, device), models, scale_factor=1, tile_size=args.tile_size)
    else:
        rlts = [process(img, model, device) for model in models]

    if args.seamless or args.mirror or args.replicate:
        rlts = [crop_seamless(rlt) for rlt in rlts]

    normal_map = rlts[0]
    roughness = rlts[1][:, :, 1]
    displacement = rlts[1][:, :, 0]

    # Save outputs
    if args.ishiiruka_texture_encoder:
        r = 255 - roughness
        g = normal_map[:, :, 1]
        b = displacement
        a = normal_map[:, :, 2]
        output = cv2.merge((b, g, r, a))
        cv2.imwrite(os.path.join(args.output, f'{base}.mat.png'), output)
    else:
        normal_name = f'{base}_Normal.png'
        cv2.imwrite(os.path.join(args.output, normal_name), normal_map)

        rough_name = f'{base}_Roughness.png'
        rough_img = 255 - roughness if args.ishiiruka else roughness
        cv2.imwrite(os.path.join(args.output, rough_name), rough_img)

        displ_name = f'{base}_Displacement.png'
        cv2.imwrite(os.path.join(args.output, displ_name), displacement)

if __name__ == "__main__":
    main()