/Anime-Colorization-Machine

User-interactive anime colorization

Primary LanguageJupyter NotebookMIT LicenseMIT

Anime-Colorization-Machine

User-interactive colorization of anime sketch image

Prerequisites

pytorch

torchvision

numpy

openCV2

PyQT5

opencv_transforms (For training) (You can simply install this by pip install opencv_transforms)

matplotlib (For training)

Results

Demo video

ex_screenshot ex_screenshot ex_screenshot ex_screenshot ex_screenshot ex_screenshot ex_screenshot ex_screenshot ex_screenshot ex_screenshot ex_screenshot ex_screenshot

Dataset

  1. Taebum Kim, "Anime Sketch Colorization Pair", https://www.kaggle.com/ktaebum/anime-sketch-colorization-pair

  2. 68K illustrations crawled from web

Usage

  1. Download model weight from here and unzip on src/model/checkpoint

  2. python main.py

  3. Load an image. The image will be automatically converted to sketch-style image.

  4. Draw scribbles and press 'colorize' button.

  5. Save the colorized image.

Training details

Parameter Value
Learning rate 2e-4
Batch size 2
Iteration 150K
Optimizer Adam
(beta1, beta2) (0.5, 0.999)
Data Augmentation RandomResizedCrop(512)
RandomHorizontalFlip()
HW CPU : Intel i5-8400
RAM : 16G
GPU : NVIDIA GTX1060 6G
Training Time About 1.37s per iteration
(About 50 hours for 150K iterations)