Here's the original paper:

low-res pdf - http://www.connellybarnes.com/work/publications/2014_portrai...

hi-res pdf - http://www.connellybarnes.com/work/publications/2014_portrai...

Thanks for the links, the two things that are really interesting to me are taking the makeup style from the example and applying it to the picture (preview the made up version of the face) and taking famous actors and morphing into something "like" them. This would be useful at creating props on screen where there are pictures from someone's "earlier" life, that could make them look much more realistic. Fascinating stuff.

The main difference of this method is doing the transfer locally - while most off-the-shelf filters only create global effects.

How does this compare to Soonmin Bae’s work from a few years ago? (I can go read the paper, but maybe you have a quick summary?)

(Another of Frédo Durand’s grad students who graduated a while back, http://people.csail.mit.edu/soonmin/photolook/)

Do you have any plans for future improvements which might reduce the really distracting color ringing artifacts you get in this current work?

Are there any photographers whose style doesn’t transfer as well? (The ones you chose all use very high contrast and harsh lighting, with an emphasis on fine structure detail.)

Both papers use multi-scale approach to capture textures in different scales. While Bae's work is global transfer, this paper uses local transfer. This paper first uses a technique called "sift flow" in computer vision to compute the correspondences between example face and input face, and then transfer the image statistics locally, based on the correspondences.

We are looking at gain maps to see if we can reduce the artifacts. Styles with hard shadow are more challenging. eg Olaf Blecker http://i1.dripimg.com/t/460000/6554/96/450324_500_8a28bd.jpg (Note that high contrast doesn't mean hard shadow)

As well as, I'm gathering from the summary, applying local filters to each image frequency band independently ("Laplacian pyramid").

Exactly. In my viewpoint, the dense correspondence is another key factor that makes this technique work.

Url changed from http://techcrunch.com/2014/05/30/mit-researchers-create-an-a..., which points to this.

More evidence of terrible mainstream scientific reporting. For some reason, bloggers love to blow every achievement and finding completely out of proportion.

quantum computing breakthrough ==> scientists invent lightsaber

new photograph analysis algorithm ==> MIT turns selfies into art

These people must have the best imaginations ever.

Ah, there it is!

The lawsuits that will happen when someone implements this are going to be interesting. Artists tend to be very protective over their styles: http://en.wikipedia.org/wiki/Leibovitz_v._Paramount_Pictures....

I don't think they can successfully copyright their style. Art is all about using and recycling other people's works. The art would not progress if there wasn't derivation. Wan't it Picasso who said "Good artists copy, great artists steal (ideas)"?

Also, artists tend to evolve over time. Few artists like to remain stuck in one style for fear of being thought as unimaginative or lazy (you could look into all the flak Ansel Adams gets). Also some like Kinkade made their style into a business and people expected a certain style so he could not get away from that.

Most successful photographers evolve their style over time --Winogrand, Friedlander, Eggleston, Atget, Frank, Moriyama, etc.

http://en.wikipedia.org/wiki/Leibovitz_v._Paramount_Pictures... <- I had to put a %2E instead of a period so HN wouldn't eat it.

the wikipedia url needs a trailing period:

http://en.wikipedia.org/wiki/Leibovitz_v._Paramount_Pictures....

[edit] that apparently HN lops off. I haven't found a way around it.

Right. If you want something not entirely dissimilar, and available right now for your photos, visit http://vsco.co or look for the VSCO Cam app on the app store.

It doesn't do based-on-image-X matchy-matching but it does good film emulation.

If you read the article it has absolutely nothing to do with global filters or film emulation.

The algorithm doesn't, but the results absolutely have.

Celebrated photographers "styles" can be two things:

a) a way of seeing (composition, sense of space, etc, etc)

b) a specific look, based on favorite film stock, preferred lighting schemes, post-processing etc.

This can immitate the second. Which you can also get, with a more manual process, from global filter/film emulation, like VSCO.

The local vs global application of the filter doesn't have as much impact in the final output. It's just a slightly more accurate (b).

It's not just that this can emulate the second. Sometimes VSCO Film is the second, or at least the foundation for the second. It's kind of swept the modern photography world by storm.

That said, to the grand-parent poster: I realize that it isn't the same thing as described in the paper! However, it is available now, and I was using phrases such as "not entirely unlike" which shouldn't exactly inspire a sense of precise equivalence...

In the beginning of the article I thought it was just some instagram-like filters with some buzzwords like "selfie" added. But when I saw the video showing that it can apply the style of one image to the other, that was where I started to find it awesome! Neat!

This is absolutely amazing. Would love to see some demo code up on GitHub.

I'm planning to release some demo code before the presentation at August :-)

Someone more familiar with tools for this and located in the US should start a bounty for submitting the first imagemagick implementation.

It's not like imagemagick has any tradition of these kind of things -- or that it's ever the first program to get those.

You'd might see something in Gimp long before there's any imagemagick implementation.

And usually those come from proprietary sources first.

Sure, but it would be useful. I for one would back a bounty.

Did the title of this link change? Didn't it originally say "MIT researchers turn selfies into art", or something like that?

In my days that stuff was just called filters.

Now you can give it an example image, and it will automatically make your image look more like the example.