search

MarkMoHR / virtual_sketching

[SIGGRAPH 2021] General Virtual Sketching Framework for Vector Line Art
https://markmohr.github.io/virtual_sketching/
Apache License 2.0
211 stars 43 forks source link

Train the neural renderer #3

Closed RongkaiZhang closed 3 years ago

RongkaiZhang commented 3 years ago

Hi Mark,

Thank you for your excellent and impressive work. I would like to train the neural renderer myself. However, if I generate a random stroke (i.e. a single fixed-width line, say wideth=3) and use L-2 loss, the trained neural renderer will output a solid dark image. I think it is due to the small difference between the dark image and the one with a single line. I cannot find the code for training the neural renderer. Could you please share small ideas or the code you used to train it?

Best regards, Rongkai

MarkMoHR commented 3 years ago

Hi,

I trained the neural renderer with the approach of Learning-to-Paint. You can try it.

Regarding your issue, I think you can also try L1 loss or Binary Cross Entropy loss. And I recommend fine-tuning on the provided model.

RongkaiZhang commented 3 years ago

Hi,

It works. Thanks.