training datasets

[dexhunter]

Hi! As @lllyasviel asked in makegirls I was wondering if @lllyasviel is willing to share the training datasets for this repo?

I have collected and preprocessed a dataset of anime faces with corresponding sketches(96x96 x2) Here is the link:

链接: https://pan.baidu.com/s/1o87HtRc 密码: yjmi

[lllyasviel]

As you wish.
400,000 images: https://nico-opendata.jp/en/seigadata/index.html
sketch maker: https://github.com/lllyasviel/sketchKeras

[dexhunter]

@lllyasviel Hi! As I checkout the niconico dataset, there are too many 'noisy'(weird cartoon, not usable) images. So do you use all of the image as input (without label?) or do you hand-pick (or with a script?) some of the images?

Edit:
Or do you use other data sets as well?

[lllyasviel]

just download the dataset in reversed order.

[dexhunter]

I don't think download in reverse order has anything to do with the image selection. Do you have any filter mechanism for the images? I doubt that you can just load all the data to the model and get the current result.

…

On Sat, Aug 19, 2017 at 12:39 AM, lllyasviel ***@***.***> wrote: just download the dataset in reversed order. — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGnAVnnV0giOcpfMF-uupnwC9RqUHmKSks5sZb5AgaJpZM4O6ig9> .

-- Kind regards, Dex

[lllyasviel]

you can download in reversed order and do not use all the dataset.
the quality can be relatively higher.

[dexhunter]

Hi! I think it's okay that you collect your own data set, but at least you should be honest about what kind of data you are using.

Clearly, this open dataset you give with so many noisy data cannot generate the illustration on README. Here is a peek to the dataset_067 & dataset_066

[lllyasviel]

sketches in readme is real sketches by human and sketches from some random google search image results' sketch from sketchKeras.
It is meaningless to use sketches in database to show the model's result.

[dexhunter]

So you hand-picked out the sketches & non-character images from data sets?

[lllyasviel]

just use noisy dataset to train and then use your favourite sketch to test your model.
good luck.

[lllyasviel]

"noisy" is important for these model including paintschainer. paintschainer even has its own way to make data noisy. noisy is fine.

[dexhunter]

Hmm...Interesting, I will take a look after the training. Thanks anyways.

[lllyasviel]

I also do not believe in the power of this noisy dataset at the begining, but not after I find this excellent project based on this dataset.
https://nico-opendata.jp/ja/casestudy/neural_style_synthesizer/index.html

[diyseguy]

Has anyone tried other datasets? I would like to run this locally (since website has been down for a while). Still waiting for approval from nico-opendata to access these images. Thank you