Coloring Grey-Scale Image: Another triumph of deep learning AI

[pandy]

Charlie Chaplin short colored by neural networks https://www.youtube.com/watch?v=_MJU8VK2PI4

Bernix

Thanks! Nice sample.
Looks like to high resolution on input and color map is not properly distributed (limitation described by paper).
Side to this somehow i start to like color version of movies (weird as i was against colorization process).
Looks like adding third layer of neural network to improve temporal stability may give quite robust movie plugin (but perhaps this is too naive expectation).

[MaverickTse]

In case you are not aware of, the Chaplin sample was created using a different set of tools and training sets. (probably not related to the Wasada team)
The underlining idea is still machine-learning however.

[Bernix]

yes. http://tinyclouds.org/colorize/

What is the ->

Download
Here is a trained TensorFlow model to play around with:
colorize-20160110.tgz.torrent 492M
This model containes the VGG16 model from Karen Simonyan and Andrew Zisserman (that I converted to TensorFlow). It is available for non-commercial use only.

For what is the 492M download good ? Is it possible to compile it ?
Sorry I am noob...

Bernix

[MaverickTse]

I don't think you can compile it... like the waifu2x model data, they are not the actual executable.

[pandy]

In case you are not aware of, the Chaplin sample was created using a different set of tools and training sets. (probably not related to the Wasada team)
The underlining idea is still machine-learning however.

Strange as seem that it suffer from same (similar issue) - unequal colour distribution.

[Bernix]

I don't think you can compile it... like the waifu2x model data, they are not the actual executable.

Thank you

Bernix

[Bernix]

Mecab added flavor option on his page.
city, sky, architecture, night and sunset or no flavor
Thank you