The (Official) PyTorch Implementation of the paper Deep Extraction of Manga Structural Lines. This project aims to extract the structural lines from 2D manga, cartoons, and illustrations.

Besides colab, There is also a gradio-like web demo. (i18n help needed!)

model_torch.py [source folder] [output folder]

Example:

model_torch.py ./pytorchTestCases/ ./pytorchResults/

You can always refer to the standardized HuggingFace pipeline to perform simple inferences. Thanks for the port and model quantization.

For manual weight download, please refer to the release section of this repo. Alternatively, you may use this link:

https://www.dropbox.com/s/y8pulix3zs73y62/erika.pth?dl=0

• Python3
• PyTorch (tested on version 1.9)
• Python-opencv

The PyTorch weights are exactly the same as the theano(!) model. I make some efforts to convert the original weights to the new model and ensure the overall error is less than 1e-3 over the image range from 0-255.

Moreover, the functional PyTorch interface allows easier fine-tuning of this model. You can also take the whole model as a sub-module for your own work (e.g., use the on-the-fly extraction of lines as a structural constraint).

I really don't want to admit it, but the legacy code looks like some artworks by a two-years old. Please refer to #5 if you have any interest.

Surprisingly, this model works quite well on color cartoons and other nijigen-like images, as long as they have clear hand-drawn lines. Simply load the image as grayscale(by default) and check out the results!

From left to right: input, sketchKeras, Anime2Sketch (considered as SOTA), Ours.

I'm glad to share some model results. Some of the images are copyrighted and I will list the original source below. Feel free to share your creaions with me in the issues section.

©IWAYUU, from the fc2 blog.

@article{li-2017-deep,
    author   = {Chengze Li and Xueting Liu and Tien-Tsin Wong},
    title    = {Deep Extraction of Manga Structural Lines},
    journal  = {ACM Transactions on Graphics (SIGGRAPH 2017 issue)},
    month    = {July},
    year     = {2017},
    volume   = {36},
    number   = {4},
    pages    = {117:1--117:12},
}

• Xueting Liu and Tien-Tsin Wong, who contributed this work
• Wenliang Wu and Ziheng Ma, who inspired me to port this great thing to PyTorch
• Toda Erika, where the project name comes from