utils.py

import numpy as np
from PIL import Image, ImageFilter
import torch
import torch.nn.functional as F
from torchvision.transforms import GaussianBlur
import math

if (not hasattr(Image, 'Resampling')):  # For older versions of Pillow
    Image.Resampling = Image

BLUR_KERNEL_SIZE = 15


def tensor_to_pil(img_tensor, batch_index=0):
    # Takes an image in a batch in the form of a tensor of shape [batch_size, channels, height, width]
    # and returns an PIL Image with the corresponding mode deduced by the number of channels

    # Take the image in the batch given by batch_index
    img_tensor = img_tensor[batch_index].unsqueeze(0)
    i = 255. * img_tensor.cpu().numpy()
    img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8).squeeze())
    return img


def pil_to_tensor(image):
    # Takes a PIL image and returns a tensor of shape [1, height, width, channels]
    image = np.array(image).astype(np.float32) / 255.0
    image = torch.from_numpy(image).unsqueeze(0)
    if len(image.shape) == 3:  # If the image is grayscale, add a channel dimension
        image = image.unsqueeze(-1)
    return image


def controlnet_hint_to_pil(tensor, batch_index=0):
    return tensor_to_pil(tensor.movedim(1, -1), batch_index)


def pil_to_controlnet_hint(img):
    return pil_to_tensor(img).movedim(-1, 1)


def crop_tensor(tensor, region):
    # Takes a tensor of shape [batch_size, height, width, channels] and crops it to the given region
    x1, y1, x2, y2 = region
    return tensor[:, y1:y2, x1:x2, :]


def resize_tensor(tensor, size, mode="nearest-exact"):
    # Takes a tensor of shape [B, C, H, W] and resizes
    # it to a shape of [B, C, size[0], size[1]] using the given mode
    return torch.nn.functional.interpolate(tensor, size=size, mode=mode)


def get_crop_region(mask, pad=0):
    # Takes a black and white PIL image in 'L' mode and returns the coordinates of the white rectangular mask region
    # Should be equivalent to the get_crop_region function from https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/modules/masking.py
    coordinates = mask.getbbox()
    if coordinates is not None:
        x1, y1, x2, y2 = coordinates
    else:
        x1, y1, x2, y2 = mask.width, mask.height, 0, 0
    # Apply padding
    x1 = max(x1 - pad, 0)
    y1 = max(y1 - pad, 0)
    x2 = min(x2 + pad, mask.width)
    y2 = min(y2 + pad, mask.height)
    return fix_crop_region((x1, y1, x2, y2), (mask.width, mask.height))


def fix_crop_region(region, image_size):
    # Remove the extra pixel added by the get_crop_region function
    image_width, image_height = image_size
    x1, y1, x2, y2 = region
    if x2 < image_width:
        x2 -= 1
    if y2 < image_height:
        y2 -= 1
    return x1, y1, x2, y2


def expand_crop(region, width, height, target_width, target_height):
    '''
    Expands a crop region to a specified target size.
    :param region: A tuple of the form (x1, y1, x2, y2) denoting the upper left and the lower right points
        of the rectangular region. Expected to have x2 > x1 and y2 > y1.
    :param width: The width of the image the crop region is from.
    :param height: The height of the image the crop region is from.
    :param target_width: The desired width of the crop region.
    :param target_height: The desired height of the crop region.
    '''
    x1, y1, x2, y2 = region
    actual_width = x2 - x1
    actual_height = y2 - y1
    # target_width = math.ceil(actual_width / 8) * 8
    # target_height = math.ceil(actual_height / 8) * 8

    # Try to expand region to the right of half the difference
    width_diff = target_width - actual_width
    x2 = min(x2 + width_diff // 2, width)
    # Expand region to the left of the difference including the pixels that could not be expanded to the right
    width_diff = target_width - (x2 - x1)
    x1 = max(x1 - width_diff, 0)
    # Try the right again
    width_diff = target_width - (x2 - x1)
    x2 = min(x2 + width_diff, width)

    # Try to expand region to the bottom of half the difference
    height_diff = target_height - actual_height
    y2 = min(y2 + height_diff // 2, height)
    # Expand region to the top of the difference including the pixels that could not be expanded to the bottom
    height_diff = target_height - (y2 - y1)
    y1 = max(y1 - height_diff, 0)
    # Try the bottom again
    height_diff = target_height - (y2 - y1)
    y2 = min(y2 + height_diff, height)

    return (x1, y1, x2, y2), (target_width, target_height)


def resize_region(region, init_size, resize_size):
    # Resize a crop so that it fits an image that was resized to the given width and height
    x1, y1, x2, y2 = region
    init_width, init_height = init_size
    resize_width, resize_height = resize_size
    x1 = math.floor(x1 * resize_width / init_width)
    x2 = math.ceil(x2 * resize_width / init_width)
    y1 = math.floor(y1 * resize_height / init_height)
    y2 = math.ceil(y2 * resize_height / init_height)
    return (x1, y1, x2, y2)


def pad_image(image, left_pad, right_pad, top_pad, bottom_pad, fill=False, blur=False):
    '''
    Pads an image with the given number of pixels on each side and fills the padding with data from the edges.
    :param image: A PIL image
    :param left_pad: The number of pixels to pad on the left side
    :param right_pad: The number of pixels to pad on the right side
    :param top_pad: The number of pixels to pad on the top side
    :param bottom_pad: The number of pixels to pad on the bottom side
    :param blur: Whether to blur the padded edges
    :return: A PIL image with size (image.width + left_pad + right_pad, image.height + top_pad + bottom_pad)
    '''
    left_edge = image.crop((0, 1, 1, image.height - 1))
    right_edge = image.crop((image.width - 1, 1, image.width, image.height - 1))
    top_edge = image.crop((1, 0, image.width - 1, 1))
    bottom_edge = image.crop((1, image.height - 1, image.width - 1, image.height))
    new_width = image.width + left_pad + right_pad
    new_height = image.height + top_pad + bottom_pad
    padded_image = Image.new(image.mode, (new_width, new_height))
    padded_image.paste(image, (left_pad, top_pad))
    if fill:
        for i in range(left_pad):
            edge = left_edge.resize(
                (1, new_height - i * (top_pad + bottom_pad) // left_pad), resample=Image.Resampling.NEAREST)
            padded_image.paste(edge, (i, i * top_pad // left_pad))
        for i in range(right_pad):
            edge = right_edge.resize(
                (1, new_height - i * (top_pad + bottom_pad) // right_pad), resample=Image.Resampling.NEAREST)
            padded_image.paste(edge, (new_width - 1 - i, i * top_pad // right_pad))
        for i in range(top_pad):
            edge = top_edge.resize(
                (new_width - i * (left_pad + right_pad) // top_pad, 1), resample=Image.Resampling.NEAREST)
            padded_image.paste(edge, (i * left_pad // top_pad, i))
        for i in range(bottom_pad):
            edge = bottom_edge.resize(
                (new_width - i * (left_pad + right_pad) // bottom_pad, 1), resample=Image.Resampling.NEAREST)
            padded_image.paste(edge, (i * left_pad // bottom_pad, new_height - 1 - i))
        if blur and not (left_pad == right_pad == top_pad == bottom_pad == 0):
            padded_image = padded_image.filter(ImageFilter.GaussianBlur(BLUR_KERNEL_SIZE))
            padded_image.paste(image, (left_pad, top_pad))
    return padded_image


def pad_image2(image, left_pad, right_pad, top_pad, bottom_pad, fill=False, blur=False):
    '''
    Pads an image with the given number of pixels on each side and fills the padding with data from the edges. 
    Faster than pad_image, but only pads with edge data in straight lines.
    :param image: A PIL image
    :param left_pad: The number of pixels to pad on the left side
    :param right_pad: The number of pixels to pad on the right side
    :param top_pad: The number of pixels to pad on the top side
    :param bottom_pad: The number of pixels to pad on the bottom side
    :param blur: Whether to blur the padded edges
    :return: A PIL image with size (image.width + left_pad + right_pad, image.height + top_pad + bottom_pad)
    '''
    left_edge = image.crop((0, 1, 1, image.height - 1))
    right_edge = image.crop((image.width - 1, 1, image.width, image.height - 1))
    top_edge = image.crop((1, 0, image.width - 1, 1))
    bottom_edge = image.crop((1, image.height - 1, image.width - 1, image.height))
    new_width = image.width + left_pad + right_pad
    new_height = image.height + top_pad + bottom_pad
    padded_image = Image.new(image.mode, (new_width, new_height))
    padded_image.paste(image, (left_pad, top_pad))
    if fill:
        if left_pad > 0:
            padded_image.paste(left_edge.resize((left_pad, new_height), resample=Image.Resampling.NEAREST), (0, 0))
        if right_pad > 0:
            padded_image.paste(right_edge.resize((right_pad, new_height),
                               resample=Image.Resampling.NEAREST), (new_width - right_pad, 0))
        if top_pad > 0:
            padded_image.paste(top_edge.resize((new_width, top_pad), resample=Image.Resampling.NEAREST), (0, 0))
        if bottom_pad > 0:
            padded_image.paste(bottom_edge.resize((new_width, bottom_pad),
                               resample=Image.Resampling.NEAREST), (0, new_height - bottom_pad))
        if blur and not (left_pad == right_pad == top_pad == bottom_pad == 0):
            padded_image = padded_image.filter(ImageFilter.GaussianBlur(BLUR_KERNEL_SIZE))
            padded_image.paste(image, (left_pad, top_pad))
    return padded_image


def pad_tensor(tensor, left_pad, right_pad, top_pad, bottom_pad, fill=False, blur=False):
    '''
    Pads an image tensor with the given number of pixels on each side and fills the padding with data from the edges.
    :param tensor: A tensor of shape [B, H, W, C]
    :param left_pad: The number of pixels to pad on the left side
    :param right_pad: The number of pixels to pad on the right side
    :param top_pad: The number of pixels to pad on the top side
    :param bottom_pad: The number of pixels to pad on the bottom side
    :param blur: Whether to blur the padded edges
    :return: A tensor of shape [B, H + top_pad + bottom_pad, W + left_pad + right_pad, C]
    '''
    batch_size, channels, height, width = tensor.shape
    h_pad = left_pad + right_pad
    v_pad = top_pad + bottom_pad
    new_width = width + h_pad
    new_height = height + v_pad

    # Create empty image
    padded = torch.zeros((batch_size, channels, new_height, new_width), dtype=tensor.dtype)

    # Copy the original image into the centor of the padded tensor
    padded[:, :, top_pad:top_pad + height, left_pad:left_pad + width] = tensor

    # Duplicate the edges of the original image into the padding
    if top_pad > 0:
        padded[:, :, :top_pad, :] = padded[:, :, top_pad:top_pad + 1, :]  # Top edge
    if bottom_pad > 0:
        padded[:, :, -bottom_pad:, :] = padded[:, :, -bottom_pad - 1:-bottom_pad, :]  # Bottom edge
    if left_pad > 0:
        padded[:, :, :, :left_pad] = padded[:, :, :, left_pad:left_pad + 1]  # Left edge
    if right_pad > 0:
        padded[:, :, :, -right_pad:] = padded[:, :, :, -right_pad - 1:-right_pad]  # Right edge

    return padded


def resize_and_pad_image(image, width, height, fill=False, blur=False):
    '''
    Resizes an image to the given width and height and pads it to the given width and height.
    :param image: A PIL image
    :param width: The width of the resized image
    :param height: The height of the resized image
    :param fill: Whether to fill the padding with data from the edges
    :param blur: Whether to blur the padded edges
    :return: A PIL image of size (width, height)
    '''
    width_ratio = width / image.width
    height_ratio = height / image.height
    if height_ratio > width_ratio:
        resize_ratio = width_ratio
    else:
        resize_ratio = height_ratio
    resize_width = round(image.width * resize_ratio)
    resize_height = round(image.height * resize_ratio)
    resized = image.resize((resize_width, resize_height), resample=Image.Resampling.LANCZOS)
    # Pad the sides of the image to get the image to the desired size that wasn't covered by the resize
    horizontal_pad = (width - resize_width) // 2
    vertical_pad = (height - resize_height) // 2
    result = pad_image2(resized, horizontal_pad, horizontal_pad, vertical_pad, vertical_pad, fill, blur)
    result = result.resize((width, height), resample=Image.Resampling.LANCZOS)
    return result, (horizontal_pad, vertical_pad)


def resize_and_pad_tensor(tensor, width, height, fill=False, blur=False):
    '''
    Resizes an image tensor to the given width and height and pads it to the given width and height.
    :param tensor: A tensor of shape [B, H, W, C]
    :param width: The width of the resized image
    :param height: The height of the resized image
    :param fill: Whether to fill the padding with data from the edges
    :param blur: Whether to blur the padded edges
    :return: A tensor of shape [B, height, width, C]
    '''
    # Resize the image to the closest size that maintains the aspect ratio
    width_ratio = width / tensor.shape[3]
    height_ratio = height / tensor.shape[2]
    if height_ratio > width_ratio:
        resize_ratio = width_ratio
    else:
        resize_ratio = height_ratio
    resize_width = round(tensor.shape[3] * resize_ratio)
    resize_height = round(tensor.shape[2] * resize_ratio)
    resized = F.interpolate(tensor, size=(resize_height, resize_width), mode='nearest-exact')
    # Pad the sides of the image to get the image to the desired size that wasn't covered by the resize
    horizontal_pad = (width - resize_width) // 2
    vertical_pad = (height - resize_height) // 2
    result = pad_tensor(resized, horizontal_pad, horizontal_pad, vertical_pad, vertical_pad, fill, blur)
    result = F.interpolate(result, size=(height, width), mode='nearest-exact')
    return result


def crop_controlnet(cond_dict, region, init_size, canvas_size, tile_size, w_pad, h_pad):
    if "control" not in cond_dict:
        return
    c = cond_dict["control"]
    controlnet = c.copy()
    cond_dict["control"] = controlnet
    while c is not None:
        # hint is shape (B, C, H, W)
        hint = controlnet.cond_hint_original
        resized_crop = resize_region(region, canvas_size, hint.shape[:-3:-1])
        hint = crop_tensor(hint.movedim(1, -1), resized_crop).movedim(-1, 1)
        hint = resize_tensor(hint, tile_size[::-1])
        controlnet.cond_hint_original = hint
        c = c.previous_controlnet
        controlnet.set_previous_controlnet(c.copy() if c is not None else None)
        controlnet = controlnet.previous_controlnet


def region_intersection(region1, region2):
    """
    Returns the coordinates of the intersection of two rectangular regions.
    :param region1: A tuple of the form (x1, y1, x2, y2) denoting the upper left and the lower right points 
        of the first rectangular region. Expected to have x2 > x1 and y2 > y1.
    :param region2: The second rectangular region with the same format as the first.
    :return: A tuple of the form (x1, y1, x2, y2) denoting the rectangular intersection. 
        None if there is no intersection.
    """
    x1, y1, x2, y2 = region1
    x1_, y1_, x2_, y2_ = region2
    x1 = max(x1, x1_)
    y1 = max(y1, y1_)
    x2 = min(x2, x2_)
    y2 = min(y2, y2_)
    if x1 >= x2 or y1 >= y2:
        return None
    return (x1, y1, x2, y2)


def crop_gligen(cond_dict, region, init_size, canvas_size, tile_size, w_pad, h_pad):
    if "gligen" not in cond_dict:
        return
    type, model, cond = cond_dict["gligen"]
    if type != "position":
        from warnings import warn
        warn(f"Unknown gligen type {type}")
        return
    cropped = []
    for c in cond:
        emb, h, w, y, x = c
        # Get the coordinates of the box in the upscaled image
        x1 = x * 8
        y1 = y * 8
        x2 = x1 + w * 8
        y2 = y1 + h * 8
        gligen_upscaled_box = resize_region((x1, y1, x2, y2), init_size, canvas_size)

        # Calculate the intersection of the gligen box and the region
        intersection = region_intersection(gligen_upscaled_box, region)
        if intersection is None:
            continue
        x1, y1, x2, y2 = intersection

        # Offset the gligen box so that the origin is at the top left of the tile region
        x1 -= region[0]
        y1 -= region[1]
        x2 -= region[0]
        y2 -= region[1]

        # Add the padding
        x1 += w_pad
        y1 += h_pad
        x2 += w_pad
        y2 += h_pad

        # Set the new position params
        h = (y2 - y1) // 8
        w = (x2 - x1) // 8
        x = x1 // 8
        y = y1 // 8
        cropped.append((emb, h, w, y, x))

    cond_dict["gligen"] = (type, model, cropped)


def crop_area(cond_dict, region, init_size, canvas_size, tile_size, w_pad, h_pad):
    if "area" not in cond_dict:
        return

    # Resize the area conditioning to the canvas size and confine it to the tile region
    h, w, y, x = cond_dict["area"]
    w, h, x, y = 8 * w, 8 * h, 8 * x, 8 * y
    x1, y1, x2, y2 = resize_region((x, y, x + w, y + h), init_size, canvas_size)
    intersection = region_intersection((x1, y1, x2, y2), region)
    if intersection is None:
        del cond_dict["area"]
        del cond_dict["strength"]
        return
    x1, y1, x2, y2 = intersection

    # Offset origin to the top left of the tile
    x1 -= region[0]
    y1 -= region[1]
    x2 -= region[0]
    y2 -= region[1]

    # Add the padding
    x1 += w_pad
    y1 += h_pad
    x2 += w_pad
    y2 += h_pad

    # Set the params for tile
    w, h = (x2 - x1) // 8, (y2 - y1) // 8
    x, y = x1 // 8, y1 // 8

    cond_dict["area"] = (h, w, y, x)


def crop_mask(cond_dict, region, init_size, canvas_size, tile_size, w_pad, h_pad):
    if "mask" not in cond_dict:
        return
    mask_tensor = cond_dict["mask"]  # (B, H, W)
    masks = []
    for i in range(mask_tensor.shape[0]):
        # Convert to PIL image
        mask = tensor_to_pil(mask_tensor, i)  # W x H

        # Resize the mask to the canvas size
        mask = mask.resize(canvas_size, Image.Resampling.BICUBIC)

        # Crop the mask to the region
        mask = mask.crop(region)

        # Add padding
        mask, _ = resize_and_pad_image(mask, tile_size[0], tile_size[1], fill=True)

        # Resize the mask to the tile size
        if tile_size != mask.size:
            mask = mask.resize(tile_size, Image.Resampling.BICUBIC)

        # Convert back to tensor
        mask = pil_to_tensor(mask)  # (1, H, W, 1)
        mask = mask.squeeze(-1)  # (1, H, W)
        masks.append(mask)

    cond_dict["mask"] = torch.cat(masks, dim=0)  # (B, H, W)


def crop_cond(cond, region, init_size, canvas_size, tile_size, w_pad=0, h_pad=0):
    cropped = []
    for emb, x in cond:
        cond_dict = x.copy()
        n = [emb, cond_dict]
        crop_controlnet(cond_dict, region, init_size, canvas_size, tile_size, w_pad, h_pad)
        crop_gligen(cond_dict, region, init_size, canvas_size, tile_size, w_pad, h_pad)
        crop_area(cond_dict, region, init_size, canvas_size, tile_size, w_pad, h_pad)
        crop_mask(cond_dict, region, init_size, canvas_size, tile_size, w_pad, h_pad)
        cropped.append(n)
    return cropped