I already know that AI can paint, by using genetic algorithms, there are already lots of works such as this and this. In addition, I also know AI can compose: Song from PI: A musically plausible network for pop music generation (genetic algorithm too).

But what I find interesting is not painting those ambiguity/abstract paintings.

The not abstract painting flow I think is (just for example):

  1. at least training AI with superman's comic

  2. give AI a very simple posture sketch of a standing human

  3. AI paint it to superman.

Currently, I don't know if there is any way/guide/thought/algorithm can teach AI to paint a superman like the comic (not abstract ones). I'd like to research this area but can't find where and how to start.

  • 2
    $\begingroup$ You may be interested in The Painting Fool, a program that was able to learn how to paint pictures. Here's an article about The Painting Fool painting people who were sitting down. I'm not sure what type of training was used though (or even if it did use any machine learning), so I won't turn this comment into an answer until I figure out the process. Doesn't seem like it's a photo filter though. $\endgroup$ Jan 9 '17 at 20:39

This paper (that was featured in another question), StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks describes some techniques similar to what you described in the question. Instead of comics, the system in the paper is trained on a database of nature photos (birds or flowers) combined with text descriptions of each photo. It will then draw an "imaginary" bird or flower based purely on a text description. It does this in two steps: it the first, it picks a posture for the depicted subject (I believe randomly...) and roughs out patches of color that match the described features of the subject in the selected posture. It then refines the sketch, again using the text description for parameters, until it appears "photo-realistic" in the style of the training photos. It does not just recreate its training images; instead, it is able to recreate features of its training images with different positions/colors.

  • $\begingroup$ Thank you for providing this amazing works !They only use two stage to generate a vivid pic, that is really amazing.Because what I thought paint a superman at least split trainning paint into eye, brown, nose, mouth , hand ,finger ,leg etc ...but they show here is a much smarter way, really shocked by the technique evolution. $\endgroup$
    – Mithril
    Jan 10 '17 at 6:58
  • $\begingroup$ @Mithril, If you like this answer, please accept it so that other people know that you like it. $\endgroup$ Jan 11 '17 at 15:01
  • $\begingroup$ @Tariq Ali This answer really a good clue for me, so I do upvote.But I need wait some time to accept an answer(also hope to see another good approch).I don't know if StackGAN can paint comic, because learn from real photo by text is different from what I ask, and I haven't make StackGAN work for now(too much knowledge I don't know.).Need time to understand how it work, and how to modify. Maybe at last would prove it can't work for my purpose , though it still an amazing project. So I need wait. $\endgroup$
    – Mithril
    Jan 12 '17 at 2:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.