Is there any published paper about this work?

[FantasyJXF]

The result looks nice. BUT, what's the dataset, how you trained the model? Hope you could share some details

[lllyasviel]

I was considering a paper 3 years ago. But finally I found it is a bit trivial and did not manage one.......

[FantasyJXF]

I'm trying to extract sketches from colored cartoon picture, but it's hard to find paired datasets, and few works to refer to.

I tried the model you provided, I suppose you trained a U-Net to do the binary classification to judge whether it's an EDGE or not? And U-Net requires less data to get precise result, is that right?

I wonder how you trained the model.

[lllyasviel]

I believe there was, maybe about at 2017, a dataset of paired illustration and aligned manually-inked line drawings. I forgot the searching keyworks but there should be one. It should be something like "pixiv line drawing" or something else...

[FantasyJXF]

pixiv Inc runs PaintsChainer. Thanks

[FantasyJXF]

I download the paired sketch and colored images from pixiv with this dataset, I'd like to ask about the training procudure, mainly the sketch img process, should I turn the sketch to binary map which has unique value 0 and 1? Loss function with BCE？