This Painter AI Fools Art Historians 39% of the Time

Crypto and PayPal links are available below. Thank you very much for your generous support!
Bitcoin: 13hhmJnLEzwXgmgJN7RB6bWVdT7WkrFAHh
Ethereum: 0x002BB163DfE89B7aD0712846F1a1E53ba6136b5A
LTC: LM8AUh5bGcNgzq6HaV1jeaJrFvmKxxgiXg

The paper "A Style-Aware Content Loss for Real-time HD Style Transfer" is available here:

We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
313V, Andrew Melnychuk, Angelos Evripiotis, Anthony Vdovitchenko, Brian Gilman, Christian Ahlin, Christoph Jadanowski, Dennis Abts, Emmanuel, Eric Haddad, Eric Martel, Esa Turkulainen, Evan Breznyik, Geronimo Moralez, John De Witt, Kjartan Olason, Lorin Atzberger, Marten Rauschenberg, Michael Albrecht, Michael Jensen, Milan Lajtoš, Morten Punnerud Engelstad, Nader Shakerin, Owen Skarpness, Raul Araújo da Silva, Rob Rowe, Robin Graham, Ryan Monsurate, Shawn Azman, Steef, Steve Messina, Sunil Kim, Thomas Krcmar, Torsten Reil, Zach Boldyga.

Károly Zsolnai-Fehér's links:

[MobyMotion]

Awesome! Always exciting to see these old techniques improved and made more robust. Also, it’s weird that I think of that original paper as old when it came out just 3 years ago haha. I can’t think of another field which moves this quickly

[lagomoof]

So can this work backwards? i.e. Can it take the knowledge of the style of an artist and determine what an original scene looked like when given a real example of the artist's work?

[willinton]

I live for that “next time” at the end of each video, they day you don’t say that I’m pulling the plug

[diamondguy]

Every day a Two Minute Papers video comes out is a good day!

[Fyloeu]

Man, that horse riding example is a genre in itself. This could be massive market for a new type of entertainment industry. And this is "just a beginning".

[Lugmillord]

The videos are especially great. Forget selfies with cat ears - the future has real time video filters.

[ProGamerGov]

When I see tables of different style transfer methods, I can't help but wonder if the authors were cherry picking or lemon picking the results, in order to make their work look better than it actually is. I've seen papers blatantly use the worst results that you can imagine, in order to represent their competition. Due to the subjective nature of this field, it's not obvious that this is happening unless you are experienced with the setup used to make the result for the research paper/project example.

[ddiva]

I love this channel, your work is incredible.
However, sometimes when there are tables with lots of images, it can be a bit confusing. Especially sense the "best" result isn't always the most vibrant, or the one with the most contrast. When the images are small it can be hard to "see" what you are talking about.
Perhaps it would be an idea to have a small colored border around certain images. This way you could say: "here, in green it's the original. The orange is the old method..." and so on.
My dream is ro study artificial intelligence, and you are my inspiration!
Keep up the good work.

[bronzekoala]

I guess many times it didn't fool the Art Historians, it was because of the composition not because of thy style.

[manictiger]

Welp, there goes the last vestige of humanity against being completely replaced by automation.
Not even artisan goods are safe.

[ibrahimi.taskiran]

After collecting all of the regenerated images, since their origin is known, one can make a model which transfers paintings into real photographic image. That would be very awesome to see what Picasso (or other painters) actually tried to paint.

[SapphFire]

The last videos looked really cool. They're relaxing to look at for some reason.

[raphirau]

Wow this is amazing! How much time it would have saved the movie "Loving Vincent"! How fast is that neural net, or how long does one render take?

[WorldOfHurt]

This would be perfect for a movie where people "dive" into paintings or something.

[huscat]

Waiting for sequel of Waking Life and Heavy Metal (1981).

[frankx]

Many years ago I dreamed of 'Picasso goggles'. Now this is possible! And if you get bored, you can just say, "Style change: Matisse".

[luisca]

I'm an animator, I've worked with Netflix and HBO to name some.. I'm working on putting together a workflow that incorporates a lot of the neural networks featured on here. I'm looking for people interested in collaborating

[GradientDude]

Small remark: The score 39% is the deception rate which is how often the neural network (trained to classify real paintings according to artists) was fooled by the proposed method.
The chance level is 1/624 ~ 0.0016 << 0.39

[c.j.]

Would be interesting to know how they came up with that 39% number. If they just asked "did an artist make this, or did an AI make this, " you might be able to get a better idea of how often the artists were just guessing. Like for a near-perfect AI, if the artists have no clue it's only going to trick them 50% of the time, on average. So 39% is probably a lot better than it first seems.

[Kram]

Can't wait for somebody to make an entire movie this way