Thanks for both these extra links -- I missed the previous HN discussion (of the paper) due to searching for Van Gogh / Picasso rather than anything else.
Whilst I agree the Washington Post article is rather lacking in technical details, it does provide a pretty nice visual idea of what the paper is about, which I think helps (and this comment thread means people can now find the paper very easily too!).
Of course, the machines don’t have a style all their own. They can only copy existing styles. True creativity is a greater challenge that machines have yet to master.
It seems to me that human artists take styles they have seen, blend it with their experiences and create their own style.
Blending multiple painting styles is just a tweak in the training....
Well, that isn't the only thing. I guess what others might complain about is the lack of composition, structured use of color, etc.
As much as paintings look similar, generated ones do not have a proper structure, the applied colors do not match the original intent and are often misplaced.
One of the obvious flaws is the reflection of the houses in the water, the painting isn't structured and the system isn't capable of discovering that kind of structure.
Paper on the other hand shows a very nice method of finding unique filters from given examples, and the results are amazing.
There's a lot of knowledge required to make a good painting besides just knowledge of other good paintings. I think it would take a long time and a huge neural net to encode those principals without human level intelligence.
What's more interesting than the article are the comments below it. The average person isn't commenting on what a cool deep learning breakthrough this is.
If AI is ever actually realized, it will occur along with a fit of conniptions about how "fake" it is. Well... at least until people fail to pass a double-blind test distinguishing between human and AI creations.
This Algorithm doesn't create a Van Gogh or a Picasso, but rather merges an artists style with another. If the algorithm were given two different photographs, what would the result be?
This is clearly pointed out in the Gandalf example. Though the resulting image may mimic Picasso's style of line and color, as a whole, it was unable to recreate the randomness and asymmetry of Picasso's style.
