|
import os |
|
import numpy as np |
|
from skimage import color, io |
|
import torch |
|
import torch.nn.functional as F |
|
from PIL import Image |
|
from models import ColorEncoder, ColorUNet |
|
from extractor.manga_panel_extractor import PanelExtractor |
|
import argparse |
|
|
|
os.environ["CUDA_VISIBLE_DEVICES"] = '0' |
|
|
|
def mkdirs(path): |
|
if not os.path.exists(path): |
|
os.makedirs(path) |
|
|
|
def Lab2RGB_out(img_lab): |
|
img_lab = img_lab.detach().cpu() |
|
img_l = img_lab[:,:1,:,:] |
|
img_ab = img_lab[:,1:,:,:] |
|
img_l = img_l + 50 |
|
pred_lab = torch.cat((img_l, img_ab), 1)[0,...].numpy() |
|
out = (np.clip(color.lab2rgb(pred_lab.transpose(1, 2, 0)), 0, 1) * 255).astype("uint8") |
|
return out |
|
|
|
def RGB2Lab(inputs): |
|
return color.rgb2lab(inputs) |
|
|
|
def Normalize(inputs): |
|
l = inputs[:, :, 0:1] |
|
ab = inputs[:, :, 1:3] |
|
l = l - 50 |
|
lab = np.concatenate((l, ab), 2) |
|
return lab.astype('float32') |
|
|
|
def numpy2tensor(inputs): |
|
out = torch.from_numpy(inputs.transpose(2,0,1)) |
|
return out |
|
|
|
def tensor2numpy(inputs): |
|
out = inputs[0,...].detach().cpu().numpy().transpose(1,2,0) |
|
return out |
|
|
|
def preprocessing(inputs): |
|
img_lab = Normalize(RGB2Lab(inputs)) |
|
img = np.array(inputs, 'float32') |
|
img = numpy2tensor(img) |
|
img_lab = numpy2tensor(img_lab) |
|
return img.unsqueeze(0), img_lab.unsqueeze(0) |
|
|
|
if __name__ == "__main__": |
|
parser = argparse.ArgumentParser(description="Colorize manga images.") |
|
parser.add_argument("-i", "--input_folder", type=str, required=True, help="Path to the input folder containing manga images.") |
|
parser.add_argument("-ckpt", "--model_checkpoint", type=str, required=True, help="Path to the model checkpoint file.") |
|
parser.add_argument("-o", "--output_folder", type=str, required=True, help="Path to the output folder where colorized images will be saved.") |
|
parser.add_argument("-ne", "--no_extractor", action="store_true", help="Do not segment the manga panels.") |
|
args = parser.parse_args() |
|
|
|
device = "cuda" |
|
|
|
ckpt = torch.load(args.model_checkpoint, map_location=lambda storage, loc: storage) |
|
|
|
colorEncoder = ColorEncoder().to(device) |
|
colorEncoder.load_state_dict(ckpt["colorEncoder"]) |
|
colorEncoder.eval() |
|
|
|
colorUNet = ColorUNet().to(device) |
|
colorUNet.load_state_dict(ckpt["colorUNet"]) |
|
colorUNet.eval() |
|
|
|
input_files = os.listdir(args.input_folder) |
|
|
|
for input_file in input_files: |
|
input_path = os.path.join(args.input_folder, input_file) |
|
|
|
if os.path.isfile(input_path): |
|
if args.no_extractor: |
|
ref_img_path = input("Please enter the path of the reference image: ") |
|
|
|
img1 = Image.open(ref_img_path).convert("RGB") |
|
width, height = img1.size |
|
img2 = Image.open(input_path).convert("RGB") |
|
|
|
img1, img1_lab = preprocessing(img1) |
|
img2, img2_lab = preprocessing(img2) |
|
|
|
img1 = img1.to(device) |
|
img1_lab = img1_lab.to(device) |
|
img2 = img2.to(device) |
|
img2_lab = img2_lab.to(device) |
|
|
|
with torch.no_grad(): |
|
img2_resize = F.interpolate(img2 / 255., size=(256, 256), mode='bilinear', recompute_scale_factor=False, align_corners=False) |
|
img1_L_resize = F.interpolate(img1_lab[:, :1, :, :] / 50., size=(256, 256), mode='bilinear', recompute_scale_factor=False, align_corners=False) |
|
|
|
color_vector = colorEncoder(img2_resize) |
|
fake_ab = colorUNet((img1_L_resize, color_vector)) |
|
fake_ab = F.interpolate(fake_ab * 110, size=(height, width), mode='bilinear', recompute_scale_factor=False, align_corners=False) |
|
|
|
fake_img = torch.cat((img1_lab[:, :1, :, :], fake_ab), 1) |
|
fake_img = Lab2RGB_out(fake_img) |
|
|
|
out_folder = os.path.join(args.output_folder, 'color') |
|
mkdirs(out_folder) |
|
out_img_path = os.path.join(out_folder, f'{os.path.splitext(input_file)[0]}_color.png') |
|
io.imsave(out_img_path, fake_img) |
|
|
|
else: |
|
panel_extractor = PanelExtractor(min_pct_panel=5, max_pct_panel=90) |
|
panels, masks, panel_masks = panel_extractor.extract(input_path) |
|
|
|
ref_img_paths = [] |
|
print("Please enter the name of the reference image in order according to the number prompts on the picture") |
|
for i in range(len(panels)): |
|
ref_img_path = input(f"{i+1}/{len(panels)} reference image:") |
|
ref_img_paths.append(ref_img_path) |
|
|
|
fake_imgs = [] |
|
for i in range(len(panels)): |
|
img1 = Image.fromarray(panels[i]).convert("RGB") |
|
width, height = img1.size |
|
img2 = Image.open(ref_img_paths[i]).convert("RGB") |
|
|
|
img1, img1_lab = preprocessing(img1) |
|
img2, img2_lab = preprocessing(img2) |
|
|
|
img1 = img1.to(device) |
|
img1_lab = img1_lab.to(device) |
|
img2 = img2.to(device) |
|
img2_lab = img2_lab.to(device) |
|
|
|
with torch.no_grad(): |
|
img2_resize = F.interpolate(img2 / 255., size=(256, 256), mode='bilinear', recompute_scale_factor=False, align_corners=False) |
|
img1_L_resize = F.interpolate(img1_lab[:,:1,:,:] / 50., size=(256, 256), mode='bilinear', recompute_scale_factor=False, align_corners=False) |
|
|
|
color_vector = colorEncoder(img2_resize) |
|
|
|
fake_ab = colorUNet((img1_L_resize, color_vector)) |
|
fake_ab = F.interpolate(fake_ab*110, size=(height, width), mode='bilinear', recompute_scale_factor=False, align_corners=False) |
|
|
|
fake_img = torch.cat((img1_lab[:,:1,:,:], fake_ab), 1) |
|
fake_img = Lab2RGB_out(fake_img) |
|
|
|
out_folder = os.path.join(args.output_folder, 'color') |
|
mkdirs(out_folder) |
|
out_img_path = os.path.join(out_folder, f'{os.path.splitext(input_file)[0]}_panel_{i}_color.png') |
|
io.imsave(out_img_path, fake_img) |
|
|
|
print(f'Colored images have been saved to: {os.path.join(args.output_folder, "color")}') |
|
|
