postprocess_lrback.py

"""Apply iterative LR back-projection to precomputed VARSR outputs.

This is intentionally independent of infer_folder.py and of all Wavelet/AdaIN
post-processing.  It consumes an original LR image y and a precomputed VARSR
image x, then applies the classical iterative back-projection update:

    x_{t+1} = clip(x_t + eta * U(y - D(x_t)), 0, 1)

D is bicubic downsampling with antialiasing, matching the LR fidelity loss in
3DSR's train_3dsr.py.  U is bicubic upsampling to the VARSR image resolution.
"""

import argparse
from pathlib import Path

import numpy as np
import torch
import torch.nn.functional as functional
from PIL import Image


DEFAULT_EXTENSIONS = ".png,.jpg,.jpeg,.JPG,.JPEG"


def parse_args():
    parser = argparse.ArgumentParser(
        description="Apply LR data-consistency back-projection to existing VARSR outputs.",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
    )
    parser.add_argument("--lr_dir", required=True, help="Folder containing the original LR images.")
    parser.add_argument("--varsr_dir", required=True, help="Folder containing raw VARSR HR outputs.")
    parser.add_argument("--output_dir", required=True, help="Folder for LR-back-projected HR images.")
    parser.add_argument("--scale", type=int, default=4, help="Expected VARSR enlargement factor.")
    parser.add_argument("--iterations", type=int, default=1, help="Number of iterative back-projection updates.")
    parser.add_argument("--step_size", type=float, default=1.0, help="Back-projection step size eta.")
    parser.add_argument(
        "--device",
        choices=("auto", "cpu", "cuda"),
        default="cpu",
        help="Compute device. CPU is the default because this postprocess is small and avoids inference GPU contention.",
    )
    parser.add_argument(
        "--gpu",
        type=int,
        default=None,
        help="Physical CUDA index used only with --device cuda/auto. Do not combine with CUDA_VISIBLE_DEVICES.",
    )
    parser.add_argument("--extensions", default=DEFAULT_EXTENSIONS, help="Comma-separated image extensions.")
    parser.add_argument("--save_ext", default=".png", help="Output extension. PNG is recommended to avoid lossy re-encoding.")
    parser.add_argument("--limit", type=int, default=0, help="Only process the first N images when greater than 0.")
    parser.add_argument("--overwrite", action="store_true", help="Overwrite existing corrected images.")
    args = parser.parse_args()

    if args.scale <= 0:
        parser.error("--scale must be positive")
    if args.iterations <= 0:
        parser.error("--iterations must be positive")
    if args.step_size <= 0:
        parser.error("--step_size must be positive")
    if args.gpu is not None and args.gpu < 0:
        parser.error("--gpu must be non-negative")
    return args


def normalize_extensions(raw_extensions):
    extensions = set()
    for raw_extension in raw_extensions.split(","):
        extension = raw_extension.strip().lower()
        if extension:
            extensions.add(extension if extension.startswith(".") else f".{extension}")
    return extensions


def iter_images(root, extensions):
    return sorted(
        path for path in root.rglob("*") if path.is_file() and path.suffix.lower() in extensions
    )


def relative_stem(path, root):
    return path.relative_to(root).with_suffix("").as_posix().lower()


def index_images(root, extensions):
    image_index = {}
    for image_path in iter_images(root, extensions):
        key = relative_stem(image_path, root)
        if key in image_index:
            raise RuntimeError(f"Duplicate relative image stem in {root}: {key}")
        image_index[key] = image_path
    return image_index


def basename_stem(path):
    return path.stem.lower()


def index_unique_basenames(root, extensions):
    """Index images by filename stem, rejecting ambiguous flat-name matches."""
    image_index = {}
    for image_path in iter_images(root, extensions):
        key = basename_stem(image_path)
        if key in image_index:
            raise RuntimeError(
                f"Duplicate image basename '{key}' in {root}: "
                f"{image_index[key]} and {image_path}. "
                "Use directories with matching relative paths instead."
            )
        image_index[key] = image_path
    return image_index


def find_lr_image(lr_images_by_relative_path, lr_images_by_basename, varsr_path, varsr_dir, lr_dir):
    """Match relative paths first, then support 3DSR-rendered nested VARSR outputs."""
    relative_key = relative_stem(varsr_path, varsr_dir)
    lr_path = lr_images_by_relative_path.get(relative_key)
    if lr_path is not None:
        return lr_path

    basename_key = basename_stem(varsr_path)
    lr_path = lr_images_by_basename.get(basename_key)
    if lr_path is not None:
        return lr_path

    raise FileNotFoundError(
        f"No LR image matching VARSR image '{varsr_path}'. Tried relative stem "
        f"'{relative_key}' and basename '{basename_key}' in {lr_dir}."
    )


def resolve_device(args):
    if args.device == "cpu":
        return torch.device("cpu")
    if not torch.cuda.is_available():
        if args.device == "cuda":
            raise RuntimeError("--device cuda was requested, but CUDA is unavailable")
        return torch.device("cpu")

    gpu = 0 if args.gpu is None else args.gpu
    if gpu >= torch.cuda.device_count():
        raise RuntimeError(
            f"Requested --gpu {gpu}, but PyTorch sees CUDA indices 0..{torch.cuda.device_count() - 1}. "
            "If CUDA_VISIBLE_DEVICES is set, remove it when selecting a physical GPU with --gpu."
        )
    return torch.device(f"cuda:{gpu}")


def pil_to_tensor(image, device):
    # np.array creates writable storage because the tensor is normalized in place below.
    tensor = torch.from_numpy(np.array(image, dtype=np.float32, copy=True))
    return tensor.permute(2, 0, 1).unsqueeze(0).div_(255.0).to(device)


def tensor_to_pil(tensor):
    array = (
        tensor.detach()
        .squeeze(0)
        .permute(1, 2, 0)
        .clamp(0.0, 1.0)
        .mul(255.0)
        .round()
        .to(torch.uint8)
        .cpu()
        .numpy()
    )
    return Image.fromarray(array, mode="RGB")


def downsample_to_lr(hr, lr_size):
    return functional.interpolate(hr, size=lr_size, mode="bicubic", antialias=True)


def lr_backproject(hr, lr, iterations, step_size):
    """Perform K classical bicubic iterative-back-projection updates."""
    corrected = hr
    for _ in range(iterations):
        residual = lr - downsample_to_lr(corrected, lr.shape[-2:])
        correction = functional.interpolate(
            residual, size=corrected.shape[-2:], mode="bicubic", antialias=True
        )
        corrected = (corrected + step_size * correction).clamp(0.0, 1.0)
    return corrected


def lr_mae(hr, lr):
    return (downsample_to_lr(hr, lr.shape[-2:]) - lr).abs().mean().item()


def output_path_for(varsr_path, varsr_dir, output_dir, save_ext):
    relative_path = varsr_path.relative_to(varsr_dir)
    suffix = save_ext if save_ext else relative_path.suffix
    if suffix and not suffix.startswith("."):
        suffix = f".{suffix}"
    return (output_dir / relative_path).with_suffix(suffix)


def main():
    args = parse_args()
    lr_dir = Path(args.lr_dir)
    varsr_dir = Path(args.varsr_dir)
    output_dir = Path(args.output_dir)
    extensions = normalize_extensions(args.extensions)

    if not lr_dir.is_dir():
        raise FileNotFoundError(f"LR directory does not exist: {lr_dir}")
    if not varsr_dir.is_dir():
        raise FileNotFoundError(f"VARSR directory does not exist: {varsr_dir}")

    lr_images = index_images(lr_dir, extensions)
    lr_images_by_basename = index_unique_basenames(lr_dir, extensions)
    varsr_images = iter_images(varsr_dir, extensions)
    if args.limit > 0:
        varsr_images = varsr_images[: args.limit]
    if not varsr_images:
        raise RuntimeError(f"No VARSR images found in {varsr_dir}")

    device = resolve_device(args)
    if device.type == "cuda":
        torch.cuda.set_device(device)
        print(f"LR back-projection device: {device} ({torch.cuda.get_device_name(device)})")
    else:
        print("LR back-projection device: cpu")
    print(
        f"Found {len(varsr_images)} VARSR image(s). Writing to {output_dir}\n"
        f"LRBP settings: scale={args.scale}, iterations={args.iterations}, step_size={args.step_size}"
    )

    processed = 0
    pre_mae_total = 0.0
    post_mae_total = 0.0
    for index, varsr_path in enumerate(varsr_images, 1):
        lr_path = find_lr_image(
            lr_images,
            lr_images_by_basename,
            varsr_path,
            varsr_dir,
            lr_dir,
        )

        output_path = output_path_for(varsr_path, varsr_dir, output_dir, args.save_ext)
        if output_path.exists() and not args.overwrite:
            print(f"[{index}/{len(varsr_images)}] skip existing {output_path}")
            continue

        with Image.open(lr_path) as lr_source, Image.open(varsr_path) as varsr_source:
            lr_image = lr_source.convert("RGB")
            varsr_image = varsr_source.convert("RGB")

        expected_size = (lr_image.width * args.scale, lr_image.height * args.scale)
        if varsr_image.size != expected_size:
            raise ValueError(
                f"Size mismatch for {varsr_path}: got {varsr_image.size}, expected {expected_size} "
                f"from LR image {lr_path} and scale {args.scale}"
            )

        with torch.inference_mode():
            lr_tensor = pil_to_tensor(lr_image, device)
            hr_tensor = pil_to_tensor(varsr_image, device)
            pre_mae = lr_mae(hr_tensor, lr_tensor)
            corrected_tensor = lr_backproject(
                hr_tensor, lr_tensor, iterations=args.iterations, step_size=args.step_size
            )
            post_mae = lr_mae(corrected_tensor, lr_tensor)
            corrected = tensor_to_pil(corrected_tensor)

        output_path.parent.mkdir(parents=True, exist_ok=True)
        save_kwargs = {"quality": 95} if output_path.suffix.lower() in {".jpg", ".jpeg"} else {}
        corrected.save(output_path, **save_kwargs)
        processed += 1
        pre_mae_total += pre_mae
        post_mae_total += post_mae
        print(
            f"[{index}/{len(varsr_images)}] {varsr_path} -> {output_path} "
            f"LR-MAE: {pre_mae:.6f} -> {post_mae:.6f}"
        )

    if processed:
        print(
            f"Completed: processed={processed}, total={len(varsr_images)}, "
            f"mean LR-MAE={pre_mae_total / processed:.6f} -> {post_mae_total / processed:.6f}"
        )
    else:
        print(f"Completed: processed=0, total={len(varsr_images)}")


if __name__ == "__main__":
    main()