The neural network was taught to turn paint from paint into artistic masterpieces

Lecture



The Neural Doodle program, made on the basis of a neural network, turns ordinary drawings or photographs into real pictures. Alex Champandard, the creator of the program, posted the source code on GitHub https://github.com/alexjc/neural-doodle, and described the construction of a neural network in a scientific work http://arxiv.org/abs/1603.01768 prepared for the nucl conference .ai Conference 2016.

The program is a Python script that generates images, taking three pictures as input parameters. A drawing or photo, a sample showing the style that the user wants to emulate in his image, as well as his “semantic map” - a schematic drawing, which can be done in any graphic editor like Microsoft Paint, are submitted to the “input”.

The neural network reads the features of the sample style, transfers them to the sketch, and at the output, Neural Doodle has a new full-fledged picture.

  The neural network was taught to turn paint from paint into artistic masterpieces

Champandar experimented on the paintings of classical artists. For example, he took the landscape of Renoir, created his simplified sketch, then added his own drawing to it in Paint.

At the exit he got a new version of his own landscape, which Renoir himself painted.

  The neural network was taught to turn paint from paint into artistic masterpieces

Just the same experiment the creator spent with a photo of Freddie Mercury, deciding to liken it to the style of the artist Mia Bergeron. Champadar took one of her paintings, made a sketch of the canvas and added a photo of a famous singer.

  The neural network was taught to turn paint from paint into artistic masterpieces   The neural network was taught to turn paint from paint into artistic masterpieces

  The neural network was taught to turn paint from paint into artistic masterpieces   The neural network was taught to turn paint from paint into artistic masterpieces

For the script to work, you will need a pre-trained neural network. Download it, as well as familiarize yourself with the instructions for setting up a local environment, on GeekTimes.

In a discussion of the program on Reddit, Shampadar said that on a modern computer (it uses MacBook Pro 2014), rendering one image of low resolution takes about an hour. His Twitter bot @DeepForger, styling photographs under the pictures of artists, is able to scale the final image up to 1080p resolution, but Neural Doodle is not capable of it yet.

Earlier in the network, Google's “Deep Dream” algorithm became popular, with which neural networks could draw pictures and imitate paintings of famous masters.


Comments


To leave a comment
If you have any suggestion, idea, thanks or comment, feel free to write. We really value feedback and are glad to hear your opinion.
To reply

The practical application of artificial intelligence

Terms: The practical application of artificial intelligence