search

nagadomi / waifu2x

Image Super-Resolution for Anime-Style Art
http://waifu2x.udp.jp/
MIT License
27.49k stars 2.71k forks source link

[train] How do you make the lines clear? #371

Closed CarbonPool closed 3 years ago

CarbonPool commented 3 years ago

test1 （use my_cunet noise3_scale2.0x_model)

test2 (use cunet noise3_scale2.0x_model)

The pre-trained cunet has clearer lines. Does it contain any special training samples?

nagadomi commented 3 years ago

I'm not sure, but it may be due to different training datasets. The main image dataset is a collection of art images, such as those posted on pixiv or danbooru (filtered by PNG format). It has clear contour lines. Anime capture image not included. Also adding some artificial data. dot images( https://github.com/nagadomi/waifu2x/issues/356#issuecomment-685111347 ), and novel images. novel image The novel image was generated by randomizing the font size and color from Aozora Bunko (public domain japanese digital library).

CarbonPool commented 3 years ago

I'm not sure, but it may be due to different training datasets. The main image dataset is a collection of art images, such as those posted on pixiv or danbooru (filtered by PNG format). It has clear contour lines. Anime capture image not included. Also adding some artificial data. dot images( #356 (comment) ), and novel images. novel image The novel image was generated by randomizing the font size and color from Aozora Bunko (public domain japanese digital library).

Thanks, I may need more resources for training.