Deep Photo Style Transfer

https://github.com/luanfujun/deep-photo-styletransfer

by mortenjorck • 9 years ago

1088 173 9 years ago

github.com

boxcardavin• 9 years ago

This is super fun and impressive, I'm going to get it going on my machine and start playing immediately.

starik36• 9 years ago

A write-up on how to get it going would be appreciated.

taherchhabra• 9 years ago

oarfish4• 9 years ago

Don't get your hopes up just yet. It's not usable in its current state, unless you have Matlab (octave doesn't work, see closed+open Github issues). There are efforts under way to port the matlab code to python, but i haven't tried the fork yet. Also, processing a single image on a K80 takes several hours, no comparison to neural style. Likewise, generating the laplacians also takes a loooong time, at least with octave. Maybe this will be accelerated in an python version.

sixQuarks• 9 years ago

What was up with the "apple" photo?

arketyp• 9 years ago

Presumably it is a demonstration of the network extracting the texture of an image, ignoring "spatiality", such that when a style/texture transfer is applied on one and the same image, the texture may very well order up differently than on the original image.

More specifically, ignoring spatiality is the effect of matching the response of the covariance matrix of the feature vectors, which was the key insight in the original style transfer paper by Gatys.

GordonS• 9 years ago

All 3 looked the same to me, but I guessed it was because I'm colour blind. It's not just that then?

stpe• 9 years ago

The colors are shifted in the output. red, yellow, green => yellow, green, red.

kreetx• 9 years ago

As in: the input and reference have the same color order, which makes it weird that the output has a different one.

pveierland• 9 years ago

The paper states that it shows swapping of texture between apples by using a manual segmentation input. A manual mask is also used to transfer the fireball texture to the perfume bottle.

teded32• 9 years ago

I think the input and the style are the same photo, which is odd. But if you look at the output; you can see that even though the colors have switched, the patterns and shapes have stayed the same. The middle apple has dots on it, but it is now green. The right apple is kinda spotty, but it is now red.

camtarn• 9 years ago

If you look at the URLs, most of them have the input in the 'examples/input/' dir, and the reference in the 'examples/style/' dir. The apple reference is in the 'examples/input/' dir. Perhaps that really does mean that the input and style were identical, and they deliberately re-used the same image URL, but to me it feels like a copy-and-paste error :)

jackcarter• 9 years ago

The "style" image (tar23.png) is a symlink to the input image[0], so I guess it's intentional. I'm still not sure why that ended up swapping the colors in the output, though.

[0] https://github.com/luanfujun/deep-photo-styletransfer/tree/m...

veli_joza• 9 years ago

Some impressive results! I would like to see some heavily stylized examples like achieving Sin City and Scanner Darkly visuals.

Was it really necessary to involve Matlab, Python and Lua?

hayksaakian• 9 years ago

now if you implemented this as a web app, i'd be sharing it with everybody

really cool samples, would be cool to upload my own photos and run it through this

Joeboy• 9 years ago

There's http://neuralstyle.com/ , based on different code that does a similar thing . It's quite GPU intensive so it takes a while and I think the free options are limited.

Sunset3• 9 years ago

I can donate free time on a GeForce 1060 6GB and a 12core XEON. If anyone wants to make a web frontend for the code in the linked repo.

jshmrsn• 9 years ago

It almost causes me goosebumps to think about how if someone asked me "imagine this clear sky, but with this red sunset over here" I could very plausibly come up with a similar result as shown in these examples.

The transfer of the flame to the blue perfume bottle looks like a very practical way to prototype marketing images.

aorth• 9 years ago

Yes! I imagine being like, "Damn, I always wanted to see Venice when it was snowing, but alas it was sunny during my short visit."

paradite• 9 years ago

Or even better, snow in a tropical country, which can never happen in real life.

wapz• 9 years ago

Unfortunately I don't think this will work with snow. It changes the colors of the objects and the snow too will be changed. That being said, being able to change the sky color and building colors looks so amazing.

gedy• 9 years ago

Wow. Our kids may be able to walk around with glasses and see the world styled in real-time as they please - perpetual sunshine, gloom, night, whatever..

gmt2027• 9 years ago

Rose tinted glasses to shelter them from the legacy of a devastated planet. How about the Black Mirror episode Men Against Fire for a real dystopian vision of this future.

daxelrod• 9 years ago

The excellent short film MORE by Mark Osborne is about similar reality-editing glasses: https://www.youtube.com/watch?v=cCeeTfsm8bk

drusepth4• 9 years ago

I actually really hope this will turn out to be true.

I'm unfamiliar with the processing power required to run this on a single image, let alone a video feed. Does anyone have any back of the napkin estimations that might show just how far away AR glasses are from this?

coldtea6• 9 years ago

I actually really hope this will turn out to be true.

Yeah, screw reality.

dsr_• 9 years ago

Yes. Go look at the thread starting here: https://news.ycombinator.com/item?id=13949385

and a figure of 20+ years comes up, assuming that you want it to be truly indistinguishable from reality.

It could be sooner with major algorithmical improvements, and the display technology itself could be 5-10 years out.

gwern2• 9 years ago

It's effectively doable now. Facebook has already shipped a realtime smartphone style transfer app using Caffe2go (no word on how much it hammers battery life, though). That was for a fixed style transfer network, but a variable style transfer only needs to tack on an additional smaller network to process the style image as well as the input, so <2x as much computation/model size.

netsharc• 9 years ago

If the computing power doesn't fit in a headgear, you could just put a massive computer (e.g. iPhone size) in your pocket and have it transmit the images using WiFi/Bluetooth to your glasses.

npgatech• 9 years ago

I think it would be amazing if Adobe incorporated some of these projects (Neural Style, etc.) as part of their "Creative Cloud" offerings...actually compute it in the cloud and return the result back to Photoshop.

ttoinou3• 9 years ago

That would be great. I was actually thinking of coding a plugin like that

phkahler3• 9 years ago

code one for darktable or gimp.

Jhsto• 9 years ago

I don't know whether you noticed it, but the paper reveals it is co-authored by people from Adobe. There very well might be intentions towards that later on.

ttoinou3• 9 years ago

That might take years

thecopy• 9 years ago

Really cool! This could prove useful for interior designing.

agumonkey3• 9 years ago

I feel overwhelmed by the domain; as an ex photoshop addict, it's already above "complex job" level. I wonder if people feel bad about AI "stealing their hobs" (not a joke).

hnarayanan3• 9 years ago

If they feel this way, they could try to learn more about this new world. It is fun and opens up so many possibilities!

Here is an (in progress) explanation I am working on aimed at beginners that might help: https://harishnarayanan.org/writing/artistic-style-transfer/

sah2ed2• 9 years ago

Nice work! Definitely bookmarking.

intoverflow22• 9 years ago

I feel Photoshop/Illustrator and 2D design in general is one of the only industries that has been pretty much stagnant for 15 years or so due to monopolies.

Compare it to 3D where artists are drowning in new tools and techniques every year, any movement in this space is a good thing.

Worth remembering there was a time when the heal brush didn't exist, a time when the clone brush didn't exist and even a time where the physical airbrush didn't exist.

agumonkey3• 9 years ago

Illustration doesn't need much tech. It has more to do with human emotions and symbols.

steve371• 9 years ago

the paper is co-authored by adobe folks

bedros• 9 years ago

awesome, wish there's no matlab requirement; can the matlab code be converted to python?

johndough2• 9 years ago

Python is Turing Complete, so yes.

prashnts• 9 years ago

While correct, I think GP is more interested in knowing if there're sufficient libraries available in Python so porting MATLAB code to Python is easy.

Looking at the source, almost everything in MATLAB can be ported though.

notthemessiah• 9 years ago

I'm sure it can, but I personally am going to go for MXNet in Julia (which is syntactically similar to MatLab) if I find the time.

http://mxnet.io/

oarfish4• 9 years ago

Someone did it already: martinbenson/deep-photo-styletransfer

VMG• 9 years ago

Premium Instagram filters as-a-service?

drodil• 9 years ago

So cool :)

dvcrn2• 9 years ago

This is super impressive and something that I didn't think would be possible without someone very skilled in photoshop going over the images.

As a photo enthusiast, I am very excited about this, but also a little worried that soon very simple apps are capable of doing the craziest of edits through the power of neural nets. Imagine the next 'deep beauty transfer', able to copy perfect skin from a model onto everyone, making everything a little more fake and less genuine.

The engineer in me now wants to understand how to build something like this from scratch but I think I'm probably lacking the math skills necessary.

peteretep2• 9 years ago

I suspect it's not going to be long until there's a rather creepy "nudity from source image" generator.

DanielBMarkham2• 9 years ago

There's your killer app.

Feed it a million or so porn images of people with all kinds of different body types. Then have it guess the closest match. Finally run this. Presto change-o! It's those x-rays glasses kids everywhere wanted for years.

I could easily imagine it being done live-motion and 3-d. Run the whole thing on a set of AR glasses.

knicholes2• 9 years ago

I'm okay with creepy. Let me know when this comes out if you find it first!

gwern2• 9 years ago

I can't help but think that DiscoGAN (https://arxiv.org/abs/1703.05192 "Learning to Discover Cross-Domain Relations with Generative Adversarial Networks", Kim et al 2017) would be perfect for this. Simply feed it a ton of photos of regular clothed people and other naked people (don't have to be the same people), and it'll learn a mapping on its own. Scale that up, refine it for a few years...

hnarayanan3• 9 years ago

Here is an (in progress) article I am working on that might help you: https://harishnarayanan.org/writing/artistic-style-transfer/

Repository of explanatory notebooks: https://github.com/hnarayanan/artistic-style-transfer

dvcrn2• 9 years ago

Oh sweet, thanks! I'm definitely bookmarking this

akshay4126• 9 years ago

This is excellent blogpost, thanks!

Edit: most comprehensive, best illustration, so far I have seen on this topic.

goldenkey2• 9 years ago

You need to make a correction. The map isnt linear with the addition of the b vector. Its an offset linear map.

Dragonai• 9 years ago

This is awesome, thanks!

Hydraulix989• 9 years ago

I think you need a little more information on the Gram matrix (maybe even ditch some of the elementary "this is gradient descent" stuff and just assume the reader knows a bit more about convnets so that you can dive deeper into the style transfer specifics -- there are plenty of other sources that cover the former already).

golergka3• 9 years ago

able to copy perfect skin from a model onto everyone, making everything a little more fake and less genuine

Everybody will do it for a couple of years, then it will get old and people will have to think of something original. Just like autotune.

coldtea6• 9 years ago

Only we still have autotune everywhere.

And 20+ years on, all cover shots etc are still photoshopped.

Kiro• 9 years ago

I'm sure a lot of the music you hear is using autotune without you realizing it. It's not some fad.

bsaunder2• 9 years ago

I didn't think would be possible without someone very skilled in photoshop going over the images

And this is similar how deep learning will likely erode the need for programming (IMHO).

Deep learning won't necessarily write programs (any more than this AI manipulates images via photoshop). Folks that say, don't worry, we can't write programs easily with AI are missing the vector. Writing programs isn't necessary for there to be widespread disruption.

Most programming is essentially hooking up I/O (of which UIs are a subset) to APIs, data stores and data manipulation. The "goal" of programming is not the code, but the functionality it provides.

AI's don't need to learn to code any more than they need to learn to use photoshop. They need to learn to provide functionality (or in this case manipulate image data).

skanderbm• 9 years ago

The are many scenarios where we still need provable code/security, i.e. health and safety related matters.

Not sure I would fully entrust a trained AI to control even an elevator door, where failure could result in bodily harm.

michaelmior2• 9 years ago

Someone still has to program the AI for the forseeable future.

tvural• 9 years ago

"AI's don't need to learn to code any more than they need to learn to use photoshop. They need to learn to provide functionality (or in this case manipulate image data)."

This is interesting. My counterpoint would be that if you rely on AI over programs you lose human-editability and determinism. So fixing a bug or adding a new feature might mean diving into some opaque model rather than adding a few lines of code. You couldn't do anything where consistency is important, like security, manipulating a database with important information, or GUI design. I think that at least protects large swaths of software development.

Even this example seems less like a replacement for Photoshop and more like a cool new feature Photoshop could add

hoschicz3• 9 years ago

_Coding_ is more about understanding what your boss/clients want and turning it into something more concrete, so it's merely a NLP problem. This will, I think, see adoption in "app/website builder" tools like Squarespace.

Then there is real programming, which IMHO will get automated in the far future.

johndough2• 9 years ago

Does anyone know how long it takes per image? The only piece of information I could find was that the run script uses 8 GPUs, which suggests that it takes a while.

jjcm3• 9 years ago

Probably a significantly long time. That said, I do wonder if you need to actually run this at the resolution of the output image. Since really this is just changing the tones in the image and not altering the details, you could probably optimize the algorithm heavily such that it works for high resolution images quickly.

As an example, I used frequency separation to split the detail layer from the tone layer in the original, high resolution stock photo of SF. From there I took the lower res (25% size) output from this script and used it as my tone layer. The results are OK: http://i.imgur.com/oakLUiE.jpg

It has the same overall tones, and some of the sharpness is preserved at a high resolution. My 30 second approach suffers from some edge glow, but I'm sure it could be greatly improved in an efficient, automated way.

oarfish4• 9 years ago

With the linked implementation (there is at least 1 fork rewriting the matlab bits in python) the three steps needed per image amount to probably over an hour on one k80. However, you probably need the -backend cudnn and -cudnn_autotune flags not to run out of memory. Downscaling the images from currently 700 to say 500 significantly speeds up the process as well. Still, it definitely takes longer than neural style.

vwbuwvb• 9 years ago

notice how the lighting in the style-transferred images isn't physically plausible according to the target images. For example, in the 5th example the house is still lit as in the original, as if by sunlight not by spotlights. Maybe that's why they've chosen to showcase examples with flat or ambiguous lighting (like the nighttime scenes or the autumn scene). The DNN doesn't model the physical reality of the scene, it doesn't get that it's a 3d world, it simply transports a high-dimensional vector (the 2d image) from one space to another. What our imaginations do is map that 2d image back into 3d before transforming it.

phkahler3• 9 years ago

And the city images that are styled into night were originally taken near dusk so the lights inside the buildings were already on but not very pronounced. It will not take a full daylight scene and decide where to turn the lights on at night.

jjcm3• 9 years ago

One interesting thing that may help the final output quality is preserving the detail layer of the original image, and then applying that to the output image. Here's my quick attempt at it: http://dev.jjcm.org/tonetransfer/

I basically just used a technique called frequency separation - it's extremely quick with the most computational part being a gaussian blur, and it allows you to separate detail from tone into two separate layers. From there I just took the detail layer of the original image and applied it to the tone layer of the output image.

omegote• 9 years ago

That's pretty good, a nice improvement. Frequency separation is one of the most common retouching techniques, but works wonders in this case as well.

epigramx• 9 years ago

Their implementation appears to lose a lot of detail in many examples. This is especially true on human faces or pictures that include human faces. Can you reproduce an improvement there?

jjcm3• 9 years ago

I didn't see any examples of human faces, can you link me to one? Happy to try it out and see the results.

woliveirajr• 9 years ago

I really like decomposing an image using wavelets (more than 2 layers, in general I use 7). You can make some good corrections on skin imperfections in the middle layers, keepin it natural (high frequencies untouched).

I'm thinking how it would be if I could you your method (or the OP) on the lower layers, which mainly control colors...

hamilyon2• 9 years ago

This is where neural network-generated images might start feeding into other neural networks. Imagine a limited dataset of tagged pictures and a vast number of styles. We could on-demand generate permutations and train network that recognize them much more accurately than they would be able to do with original limited dataset

gtani2• 9 years ago

That's funny, i was just going over OS X makefile with somebody yesterday and i came here to look for the thunderbolt external GPU adapter.

The repo author is super responsive if you're RAM constrained:

https://www.reddit.com/r/CUDA/comments/61gpzd/make_error_whe...

and https://github.com/luanfujun/deep-photo-styletransfer/issues...

________________

gtani2• 9 years ago

oh, that similar project repo seems to be the inspiration/genesis of this thread's repo.

Also that repo's author did the Stanford lecture comparing theano, tensorflow, torch which is very worth digging out of the various archives

kowdermeister• 9 years ago

Now imagine what we could do if it were computed in real time, 60FPS, 100x resolution/detail, embeded in AR or contact lenses.

Yep, sci-fi. But I often imagine that sometime soon I'll have a server farm in my pocket (I know I already has via the cloud, but it's a whole new game if you can do the computation on-site, low latency)

_m8fo• 9 years ago

Wow, this is ridiculously impressive. I wonder what the audio equivalent of this would be.

Anyone know how long this takes per image?

visarga• 9 years ago

I wonder what the audio equivalent of this would be.

Voice transfer. Imagine your words in the voice of any celeb.

zild3d• 9 years ago

or timbre transfer in general. Drum recording on a cheap beginner drumset -> high end Tama kit

isseu• 9 years ago

Same rhythm with a different instrument? The problem is that audio is very dense for this (now)

guepe• 9 years ago

Funny I saw a very similar attempt at Fitchburg Art Museum (MA) over the week-end. I seem to recall this was an MIT-project: is this related in some way ? It was coming with a rather complex (at least at first sight) interface allowing quite a bit of operations/transformations, which does not appear here, so this might be a separate attempt.

Looks like this thing is "in the air".

markab21• 9 years ago

It seems this could have application in conceptual interior design. Think room makeovers.

wyldfire2• 9 years ago

Wow -- lua, Python, matlab and CUDA.

I see now that the image segmentation is probably the key element to getting these stunning results. Other style transfers I've seen took abstract elements from the donor picture, but this really captures the donor and transforms the recipient significantly.

vwcx• 9 years ago

I am a professional photo editor for a major magazine. Despite what the title sounds like, I spend much of my day sending photographers on assignment and sourcing images rather than manipulating files. I often wonder how my career will evolve in ten years; the writing is on the wall for my job, given the state of the media market.

Looking at this, I now see a future as a forensic imaging specialist. At some point, the algorithm will get pretty damn good, and it will be my job to look for tells -- cracks in the generated image where reality doesn't quite line up. The question will be whether I am seeking out these abnormalities to cover them up or to call them out.

lucidrains6• 9 years ago

you may need to train some powerful discriminator networks to do that job instead of relying on the one in your head soon..

hoschicz3• 9 years ago

"forensic imaging specialists" already exist in modern GAN networks and are called discriminators. They examine how likely is it that they would be from a human and then provide feedback to the generator network on what a human would do differently.

saurabh20n• 9 years ago

Very interesting. I work in AI. Curious about trends in your industry. Please ping me at the email address in my profile.

rl3• 9 years ago

Someone should create an entire film using this technique. Shoot two films in parallel with similar scenes, then see what happens when they're blended.

Give it a creepy and/or surrealist plot so the ethereal-looking output suits the film. Perhaps the visuals could be the result of viewing the world through a robot or AI with imperfect cognition. Would be an interesting twist on the old unreliable narrator trope.

draw_down• 9 years ago

This is so cool, such an interesting and great idea. Really impressive and well-done.

Cofike• 9 years ago

Holy cow, this is amazing. The results are way better than I would ever expect as possible.

pseudobry• 9 years ago

For anyone interested in a sci-fi book about how a society filled with this kind of technology might work, I recommend reading Rainbow's End by Vernor Vinge.

https://en.wikipedia.org/wiki/Rainbows_End

double051• 9 years ago

Those results look amazing! I'm curious how well it holds up at higher resolutions and closer inspection.

Also, there doesn't seem to be any mention of the distribution license for the project. Would any of the maintainers be able to add a license to the repository? Thanks!

oarfish4• 9 years ago

there is an issue discussing that; the problem is that the work was done at adobe labs, so they will determine the license.

rcthompson2• 9 years ago

This is really cool, but the only thing I can think right now, the question that's eating away at my soul, is: why did it color-cycle the apples?

isoprophlex• 9 years ago

There are also semantic masks applied (can be found in the github repo), I think that's how the Colorado swapping is done.

rcthompson2• 9 years ago

Ah, I see. They're right here:

Input mask: https://github.com/luanfujun/deep-photo-styletransfer/blob/m... Output mask: https://github.com/luanfujun/deep-photo-styletransfer/blob/m...

So, presumably these masks were drawn manually. The three different colors in the input mask tell it to train 3 different models from the masked regions, and then the same three colors in the output mask tell it where to apply each trained model. So it trains a model for each apple and then applies that model to the next apple over. It's not just randomly swapping colors on the apples.

alanbernstein• 9 years ago

I'm afraid the answer, like the answers to so many other soul-eating questions, is that it's random.

jtraffic• 9 years ago

I wondered what was new about this particular implementation (since I've seen several others, notably the one this code is partially based on, from Justin Johnson). From the paper:

"Our contribution is to constrain the transformation from the input to the output to be locally affine in colorspace, and to express this constraint as a custom CNN layer through which we can backpropagate. We show that this approach successfully suppresses distortion and yields satisfying photorealistic style transfers in a broad variety of scenarios, including transfer of the time of day, weather, season, and artistic edits"

I was skeptical at first (even posted then deleted a sort of negative comment), but now that I read this, I see the value. The images are much more crisp and distortion free.

folli• 9 years ago

Anyone's got a link to a good write-up on how such a style transfer works on a high level?

mistercow• 9 years ago

Jeez, how many style transfer papers have there been in the last year? It's awesome, but what an odd thing to become its own subfield.

a7958479• 9 years ago

A ball

Deep Photo Style Transfer

Related Stories