esrgan_model.py

"""
Taken from https://github.com/finegrain-ai/refiners
Modified from https://github.com/philz1337x/clarity-upscaler
which is a copy of https://github.com/AUTOMATIC1111/stable-diffusion-webui
which is a copy of https://github.com/victorca25/iNNfer
which is a copy of https://github.com/xinntao/ESRGAN
"""

import math
from pathlib import Path
from typing import NamedTuple

import numpy as np
import numpy.typing as npt
import torch
import torch.nn as nn
from PIL import Image
from huggingface_hub import hf_hub_download


def conv_block(in_nc: int, out_nc: int) -> nn.Sequential:
    return nn.Sequential(
        nn.Conv2d(in_nc, out_nc, kernel_size=3, padding=1),
        nn.LeakyReLU(negative_slope=0.2, inplace=True),
    )


class ResidualDenseBlock_5C(nn.Module):
    """
    Residual Dense Block
    The core module of paper: (Residual Dense Network for Image Super-Resolution, CVPR 18)
    Modified options that can be used:
        - "Partial Convolution based Padding" arXiv:1811.11718
        - "Spectral normalization" arXiv:1802.05957
        - "ICASSP 2020 - ESRGAN+ : Further Improving ESRGAN" N. C.
            {Rakotonirina} and A. {Rasoanaivo}
    """

    def __init__(self, nf: int = 64, gc: int = 32) -> None:
        super().__init__()  # type: ignore[reportUnknownMemberType]

        self.conv1 = conv_block(nf, gc)
        self.conv2 = conv_block(nf + gc, gc)
        self.conv3 = conv_block(nf + 2 * gc, gc)
        self.conv4 = conv_block(nf + 3 * gc, gc)
        # Wrapped in Sequential because of key in state dict.
        self.conv5 = nn.Sequential(nn.Conv2d(nf + 4 * gc, nf, kernel_size=3, padding=1))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x1 = self.conv1(x)
        x2 = self.conv2(torch.cat((x, x1), 1))
        x3 = self.conv3(torch.cat((x, x1, x2), 1))
        x4 = self.conv4(torch.cat((x, x1, x2, x3), 1))
        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
        return x5 * 0.2 + x


class RRDB(nn.Module):
    """
    Residual in Residual Dense Block
    (ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks)
    """

    def __init__(self, nf: int) -> None:
        super().__init__()  # type: ignore[reportUnknownMemberType]
        self.RDB1 = ResidualDenseBlock_5C(nf)
        self.RDB2 = ResidualDenseBlock_5C(nf)
        self.RDB3 = ResidualDenseBlock_5C(nf)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        out = self.RDB1(x)
        out = self.RDB2(out)
        out = self.RDB3(out)
        return out * 0.2 + x


class Upsample2x(nn.Module):
    """Upsample 2x."""

    def __init__(self) -> None:
        super().__init__()  # type: ignore[reportUnknownMemberType]

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return nn.functional.interpolate(x, scale_factor=2.0)  # type: ignore


class ShortcutBlock(nn.Module):
    """Elementwise sum the output of a submodule to its input"""

    def __init__(self, submodule: nn.Module) -> None:
        super().__init__()  # type: ignore[reportUnknownMemberType]
        self.sub = submodule

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return x + self.sub(x)


class RRDBNet(nn.Module):
    def __init__(self, in_nc: int, out_nc: int, nf: int, nb: int) -> None:
        super().__init__()  # type: ignore[reportUnknownMemberType]
        assert in_nc % 4 != 0  # in_nc is 3

        self.model = nn.Sequential(
            nn.Conv2d(in_nc, nf, kernel_size=3, padding=1),
            ShortcutBlock(
                nn.Sequential(
                    *(RRDB(nf) for _ in range(nb)),
                    nn.Conv2d(nf, nf, kernel_size=3, padding=1),
                )
            ),
            Upsample2x(),
            nn.Conv2d(nf, nf, kernel_size=3, padding=1),
            nn.LeakyReLU(negative_slope=0.2, inplace=True),
            Upsample2x(),
            nn.Conv2d(nf, nf, kernel_size=3, padding=1),
            nn.LeakyReLU(negative_slope=0.2, inplace=True),
            nn.Conv2d(nf, nf, kernel_size=3, padding=1),
            nn.LeakyReLU(negative_slope=0.2, inplace=True),
            nn.Conv2d(nf, out_nc, kernel_size=3, padding=1),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.model(x)


def infer_params(state_dict: dict[str, torch.Tensor]) -> tuple[int, int, int, int, int]:
    # this code is adapted from https://github.com/victorca25/iNNfer
    scale2x = 0
    scalemin = 6
    n_uplayer = 0
    out_nc = 0
    nb = 0

    for block in list(state_dict):
        parts = block.split(".")
        n_parts = len(parts)
        if n_parts == 5 and parts[2] == "sub":
            nb = int(parts[3])
        elif n_parts == 3:
            part_num = int(parts[1])
            if part_num > scalemin and parts[0] == "model" and parts[2] == "weight":
                scale2x += 1
            if part_num > n_uplayer:
                n_uplayer = part_num
                out_nc = state_dict[block].shape[0]
        assert "conv1x1" not in block  # no ESRGANPlus

    nf = state_dict["model.0.weight"].shape[0]
    in_nc = state_dict["model.0.weight"].shape[1]
    scale = 2**scale2x

    assert out_nc > 0
    assert nb > 0

    return in_nc, out_nc, nf, nb, scale  # 3, 3, 64, 23, 4


Tile = tuple[int, int, Image.Image]
Tiles = list[tuple[int, int, list[Tile]]]


# https://github.com/philz1337x/clarity-upscaler/blob/e0cd797198d1e0e745400c04d8d1b98ae508c73b/modules/images.py#L64
class Grid(NamedTuple):
    tiles: Tiles
    tile_w: int
    tile_h: int
    image_w: int
    image_h: int
    overlap: int


# adapted from https://github.com/philz1337x/clarity-upscaler/blob/e0cd797198d1e0e745400c04d8d1b98ae508c73b/modules/images.py#L67
def split_grid(image: Image.Image, tile_w: int = 512, tile_h: int = 512, overlap: int = 64) -> Grid:
    w = image.width
    h = image.height

    non_overlap_width = tile_w - overlap
    non_overlap_height = tile_h - overlap

    cols = max(1, math.ceil((w - overlap) / non_overlap_width))
    rows = max(1, math.ceil((h - overlap) / non_overlap_height))

    dx = (w - tile_w) / (cols - 1) if cols > 1 else 0
    dy = (h - tile_h) / (rows - 1) if rows > 1 else 0

    grid = Grid([], tile_w, tile_h, w, h, overlap)
    for row in range(rows):
        row_images: list[Tile] = []
        y1 = max(min(int(row * dy), h - tile_h), 0)
        y2 = min(y1 + tile_h, h)
        for col in range(cols):
            x1 = max(min(int(col * dx), w - tile_w), 0)
            x2 = min(x1 + tile_w, w)
            tile = image.crop((x1, y1, x2, y2))
            row_images.append((x1, tile_w, tile))
        grid.tiles.append((y1, tile_h, row_images))

    return grid


# https://github.com/philz1337x/clarity-upscaler/blob/e0cd797198d1e0e745400c04d8d1b98ae508c73b/modules/images.py#L104
def combine_grid(grid: Grid):
    def make_mask_image(r: npt.NDArray[np.float32]) -> Image.Image:
        r = r * 255 / grid.overlap
        return Image.fromarray(r.astype(np.uint8), "L")

    mask_w = make_mask_image(
        np.arange(grid.overlap, dtype=np.float32).reshape((1, grid.overlap)).repeat(grid.tile_h, axis=0)
    )
    mask_h = make_mask_image(
        np.arange(grid.overlap, dtype=np.float32).reshape((grid.overlap, 1)).repeat(grid.image_w, axis=1)
    )

    combined_image = Image.new("RGB", (grid.image_w, grid.image_h))
    for y, h, row in grid.tiles:
        combined_row = Image.new("RGB", (grid.image_w, h))
        for x, w, tile in row:
            if x == 0:
                combined_row.paste(tile, (0, 0))
                continue

            combined_row.paste(tile.crop((0, 0, grid.overlap, h)), (x, 0), mask=mask_w)
            combined_row.paste(tile.crop((grid.overlap, 0, w, h)), (x + grid.overlap, 0))

        if y == 0:
            combined_image.paste(combined_row, (0, 0))
            continue

        combined_image.paste(
            combined_row.crop((0, 0, combined_row.width, grid.overlap)),
            (0, y),
            mask=mask_h,
        )
        combined_image.paste(
            combined_row.crop((0, grid.overlap, combined_row.width, h)),
            (0, y + grid.overlap),
        )

    return combined_image


class UpscalerESRGAN:
    def __init__(self, model_path: Path, device: torch.device, dtype: torch.dtype):
        self.model_path = model_path
        self.device = device
        self.model = self.load_model(model_path)
        self.to(device, dtype)

    def __call__(self, img: Image.Image) -> Image.Image:
        return self.upscale_without_tiling(img)

    def to(self, device: torch.device, dtype: torch.dtype):
        self.device = device
        self.dtype = dtype
        self.model.to(device=device, dtype=dtype)

    def load_model(self, path: Path) -> RRDBNet:
        filename = path
        state_dict: dict[str, torch.Tensor] = torch.load(filename, weights_only=True, map_location=self.device)  # type: ignore
        in_nc, out_nc, nf, nb, upscale = infer_params(state_dict)
        assert upscale == 4, "Only 4x upscaling is supported"
        model = RRDBNet(in_nc=in_nc, out_nc=out_nc, nf=nf, nb=nb)
        model.load_state_dict(state_dict)
        model.eval()

        return model

    def upscale_without_tiling(self, img: Image.Image) -> Image.Image:
        img_np = np.array(img)
        img_np = img_np[:, :, ::-1]
        img_np = np.ascontiguousarray(np.transpose(img_np, (2, 0, 1))) / 255
        img_t = torch.from_numpy(img_np).float()  # type: ignore
        img_t = img_t.unsqueeze(0).to(device=self.device, dtype=self.dtype)
        with torch.no_grad():
            output = self.model(img_t)
        output = output.squeeze().float().cpu().clamp_(0, 1).numpy()
        output = 255.0 * np.moveaxis(output, 0, 2)
        output = output.astype(np.uint8)
        output = output[:, :, ::-1]
        return Image.fromarray(output, "RGB")

    # https://github.com/philz1337x/clarity-upscaler/blob/e0cd797198d1e0e745400c04d8d1b98ae508c73b/modules/esrgan_model.py#L208
    def upscale_with_tiling(self, img: Image.Image) -> Image.Image:
        img = img.convert("RGB")
        grid = split_grid(img)
        newtiles: Tiles = []
        scale_factor: int = 1

        for y, h, row in grid.tiles:
            newrow: list[Tile] = []
            for tiledata in row:
                x, w, tile = tiledata
                output = self.upscale_without_tiling(tile)
                scale_factor = output.width // tile.width
                newrow.append((x * scale_factor, w * scale_factor, output))
            newtiles.append((y * scale_factor, h * scale_factor, newrow))

        newgrid = Grid(
            newtiles,
            grid.tile_w * scale_factor,
            grid.tile_h * scale_factor,
            grid.image_w * scale_factor,
            grid.image_h * scale_factor,
            grid.overlap * scale_factor,
        )
        output = combine_grid(newgrid)
        return output