melMass-comfy_mtb/nodes/deep_bump.py

import tempfile
from pathlib import Path

import numpy as np
import onnxruntime as ort
import torch
from PIL import Image

from ..errors import ModelNotFound
from ..log import mklog
from ..utils import get_model_path, tensor2pil, tiles_infer, tiles_merge, tiles_split

# Disable MS telemetry
ort.disable_telemetry_events()
log = mklog(__name__)


# - COLOR to NORMALS
def color_to_normals(color_img, overlap, progress_callback, save_temp=False):
    """Computes a normal map from the given color map. 'color_img' must be a numpy array
    in C,H,W format (with C as RGB). 'overlap' must be one of 'SMALL', 'MEDIUM', 'LARGE'.
    """
    temp_dir = Path(tempfile.mkdtemp()) if save_temp else None

    # Remove alpha & convert to grayscale
    img = np.mean(color_img[:3], axis=0, keepdims=True)

    if temp_dir:
        Image.fromarray((img[0] * 255).astype(np.uint8)).save(
            temp_dir / "grayscale_img.png"
        )

    log.debug(
        f"Converting color image to grayscale by taking the mean over color channels: {img.shape}"
    )

    # Split image in tiles
    log.debug("DeepBump Color → Normals : tilling")
    tile_size = 256
    overlaps = {
        "SMALL": tile_size // 6,
        "MEDIUM": tile_size // 4,
        "LARGE": tile_size // 2,
    }
    stride_size = tile_size - overlaps[overlap]
    tiles, paddings = tiles_split(
        img, (tile_size, tile_size), (stride_size, stride_size)
    )
    if temp_dir:
        for i, tile in enumerate(tiles):
            Image.fromarray((tile[0] * 255).astype(np.uint8)).save(
                temp_dir / f"tile_{i}.png"
            )

    # Load model
    log.debug("DeepBump Color → Normals : loading model")
    model = get_model_path("deepbump", "deepbump256.onnx")
    if not model or not model.exists():
        raise ModelNotFound(f"deepbump ({model})")

    ort_session = ort.InferenceSession(model)

    # Predict normal map for each tile
    log.debug("DeepBump Color → Normals : generating")
    pred_tiles = tiles_infer(tiles, ort_session, progress_callback=progress_callback)

    if temp_dir:
        for i, pred_tile in enumerate(pred_tiles):
            Image.fromarray((pred_tile.transpose(1, 2, 0) * 255).astype(np.uint8)).save(
                temp_dir / f"pred_tile_{i}.png"
            )

    # Merge tiles
    log.debug("DeepBump Color → Normals : merging")
    pred_img = tiles_merge(
        pred_tiles,
        (stride_size, stride_size),
        (3, img.shape[1], img.shape[2]),
        paddings,
    )

    if temp_dir:
        Image.fromarray((pred_img.transpose(1, 2, 0) * 255).astype(np.uint8)).save(
            temp_dir / "merged_img.png"
        )

    # Normalize each pixel to unit vector
    pred_img = normalize(pred_img)

    if temp_dir:
        Image.fromarray((pred_img.transpose(1, 2, 0) * 255).astype(np.uint8)).save(
            temp_dir / "final_img.png"
        )

        log.debug(f"Debug images saved in {temp_dir}")

    return pred_img


# - NORMALS to CURVATURE
def conv_1d(array, kernel_1d):
    """Performs row by row 1D convolutions of the given 2D image with the given 1D kernel."""

    # Input kernel length must be odd
    k_l = len(kernel_1d)
    assert k_l % 2 != 0
    # Convolution is repeat-padded
    extended = np.pad(array, k_l // 2, mode="wrap")
    # Output has same size as input (padded, valid-mode convolution)
    output = np.empty(array.shape)
    for i in range(array.shape[0]):
        output[i] = np.convolve(extended[i + (k_l // 2)], kernel_1d, mode="valid")

    return output * -1


def gaussian_kernel(length, sigma):
    """Returns a 1D gaussian kernel of size 'length'."""

    space = np.linspace(-(length - 1) / 2, (length - 1) / 2, length)
    kernel = np.exp(-0.5 * np.square(space) / np.square(sigma))
    return kernel / np.sum(kernel)


def normalize(np_array):
    """Normalize all elements of the given numpy array to [0,1]"""

    return (np_array - np.min(np_array)) / (np.max(np_array) - np.min(np_array))


def normals_to_curvature(normals_img, blur_radius, progress_callback):
    """Computes a curvature map from the given normal map. 'normals_img' must be a numpy array
    in C,H,W format (with C as RGB). 'blur_radius' must be one of 'SMALLEST', 'SMALLER', 'SMALL',
    'MEDIUM', 'LARGE', 'LARGER', 'LARGEST'."""

    # Convolutions on normal map red & green channels
    if progress_callback is not None:
        progress_callback(0, 4)
    diff_kernel = np.array([-1, 0, 1])
    h_conv = conv_1d(normals_img[0, :, :], diff_kernel)
    if progress_callback is not None:
        progress_callback(1, 4)
    v_conv = conv_1d(-1 * normals_img[1, :, :].T, diff_kernel).T
    if progress_callback is not None:
        progress_callback(2, 4)

    # Sum detected edges
    edges_conv = h_conv + v_conv

    # Blur radius size is proportional to img sizes
    blur_factors = {
        "SMALLEST": 1 / 256,
        "SMALLER": 1 / 128,
        "SMALL": 1 / 64,
        "MEDIUM": 1 / 32,
        "LARGE": 1 / 16,
        "LARGER": 1 / 8,
        "LARGEST": 1 / 4,
    }
    assert blur_radius in blur_factors
    blur_radius_px = int(np.mean(normals_img.shape[1:3]) * blur_factors[blur_radius])

    # If blur radius too small, do not blur
    if blur_radius_px < 2:
        edges_conv = normalize(edges_conv)
        return np.stack([edges_conv, edges_conv, edges_conv])

    # Make sure blur kernel length is odd
    if blur_radius_px % 2 == 0:
        blur_radius_px += 1

    # Blur curvature with separated convolutions
    sigma = blur_radius_px // 8
    if sigma == 0:
        sigma = 1
    g_kernel = gaussian_kernel(blur_radius_px, sigma)
    h_blur = conv_1d(edges_conv, g_kernel)
    if progress_callback is not None:
        progress_callback(3, 4)
    v_blur = conv_1d(h_blur.T, g_kernel).T
    if progress_callback is not None:
        progress_callback(4, 4)

    # Normalize to [0,1]
    curvature = normalize(v_blur)

    # Expand single channel the three channels (RGB)
    return np.stack([curvature, curvature, curvature])


# - NORMALS to HEIGHT
def normals_to_grad(normals_img):
    return (normals_img[0] - 0.5) * 2, (normals_img[1] - 0.5) * 2


def copy_flip(grad_x, grad_y):
    """Concat 4 flipped copies of input gradients (makes them wrap).
    Output is twice bigger in both dimensions."""

    grad_x_top = np.hstack([grad_x, -np.flip(grad_x, axis=1)])
    grad_x_bottom = np.hstack([np.flip(grad_x, axis=0), -np.flip(grad_x)])
    new_grad_x = np.vstack([grad_x_top, grad_x_bottom])

    grad_y_top = np.hstack([grad_y, np.flip(grad_y, axis=1)])
    grad_y_bottom = np.hstack([-np.flip(grad_y, axis=0), -np.flip(grad_y)])
    new_grad_y = np.vstack([grad_y_top, grad_y_bottom])

    return new_grad_x, new_grad_y


def frankot_chellappa(grad_x, grad_y, progress_callback=None):
    """Frankot-Chellappa depth-from-gradient algorithm."""

    if progress_callback is not None:
        progress_callback(0, 3)

    rows, cols = grad_x.shape

    rows_scale = (np.arange(rows) - (rows // 2 + 1)) / (rows - rows % 2)
    cols_scale = (np.arange(cols) - (cols // 2 + 1)) / (cols - cols % 2)

    u_grid, v_grid = np.meshgrid(cols_scale, rows_scale)

    u_grid = np.fft.ifftshift(u_grid)
    v_grid = np.fft.ifftshift(v_grid)

    if progress_callback is not None:
        progress_callback(1, 3)

    grad_x_F = np.fft.fft2(grad_x)
    grad_y_F = np.fft.fft2(grad_y)

    if progress_callback is not None:
        progress_callback(2, 3)

    nominator = (-1j * u_grid * grad_x_F) + (-1j * v_grid * grad_y_F)
    denominator = (u_grid**2) + (v_grid**2) + 1e-16

    Z_F = nominator / denominator
    Z_F[0, 0] = 0.0

    Z = np.real(np.fft.ifft2(Z_F))

    if progress_callback is not None:
        progress_callback(3, 3)

    return (Z - np.min(Z)) / (np.max(Z) - np.min(Z))


def normals_to_height(normals_img, seamless, progress_callback):
    """Computes a height map from the given normal map. 'normals_img' must be a numpy array
    in C,H,W format (with C as RGB). 'seamless' is a bool that should indicates if 'normals_img'
    is seamless."""

    # Flip height axis
    flip_img = np.flip(normals_img, axis=1)

    # Get gradients from normal map
    grad_x, grad_y = normals_to_grad(flip_img)
    grad_x = np.flip(grad_x, axis=0)
    grad_y = np.flip(grad_y, axis=0)

    # If non-seamless chosen, expand gradients
    if not seamless:
        grad_x, grad_y = copy_flip(grad_x, grad_y)

    # Compute height
    pred_img = frankot_chellappa(-grad_x, grad_y, progress_callback=progress_callback)

    # Cut to valid part if gradients were expanded
    if not seamless:
        height, width = normals_img.shape[1], normals_img.shape[2]
        pred_img = pred_img[:height, :width]

    # Expand single channel the three channels (RGB)
    return np.stack([pred_img, pred_img, pred_img])


# - ADDON
class DeepBump:
    """Normal & height maps generation from single pictures"""

    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "image": ("IMAGE",),
                "mode": (
                    ["Color to Normals", "Normals to Curvature", "Normals to Height"],
                ),
                "color_to_normals_overlap": (["SMALL", "MEDIUM", "LARGE"],),
                "normals_to_curvature_blur_radius": (
                    [
                        "SMALLEST",
                        "SMALLER",
                        "SMALL",
                        "MEDIUM",
                        "LARGE",
                        "LARGER",
                        "LARGEST",
                    ],
                ),
                "normals_to_height_seamless": ("BOOLEAN", {"default": True}),
            },
        }

    RETURN_TYPES = ("IMAGE",)
    FUNCTION = "apply"

    CATEGORY = "mtb/textures"

    def apply(
        self,
        image,
        mode="Color to Normals",
        color_to_normals_overlap="SMALL",
        normals_to_curvature_blur_radius="SMALL",
        normals_to_height_seamless=True,
    ):
        images = tensor2pil(image)
        out_images = []

        for image in images:
            log.debug(f"Input image shape: {image}")

            in_img = np.transpose(image, (2, 0, 1)) / 255
            log.debug(f"transposed for deep image shape: {in_img.shape}")
            out_img = None

            # Apply processing
            if mode == "Color to Normals":
                out_img = color_to_normals(in_img, color_to_normals_overlap, None)
            if mode == "Normals to Curvature":
                out_img = normals_to_curvature(
                    in_img, normals_to_curvature_blur_radius, None
                )
            if mode == "Normals to Height":
                out_img = normals_to_height(in_img, normals_to_height_seamless, None)

            if out_img is not None:
                log.debug(f"Output image shape: {out_img.shape}")
                out_images.append(
                    torch.from_numpy(
                        np.transpose(out_img, (1, 2, 0)).astype(np.float32)
                    ).unsqueeze(0)
                )
            else:
                log.error("No out img... This should not happen")
        for outi in out_images:
            log.debug(f"Shape fed to utils: {outi.shape}")
        return (torch.cat(out_images, dim=0),)


__nodes__ = [DeepBump]