Hacker News new | past | comments | ask | show | jobs | submit login
Show HN: Neural network turns photos into a specific style of painting (instapainting.com)
75 points by chrischen on Sept 2, 2015 | hide | past | favorite | 41 comments




> I'd like to understand this algorithm well enough to generalize it to other media (mainly thinking about sound right now)

A DeepDream for audio would be very interesting.


It'd probably be a lot better for usable results, since you can easily restrict it to certain musical scales.


In the process of switching to this implementation.


Can't switch until nvidia approves my CUDA account.


Why do you need a CUDA account? You don't need one to install CUDA.


But you do need one for cuDNN, which is not required for Caffe, but speeds it up.


Yes getting out-of-memory issues without cudnn.


Might want to take a look at: https://github.com/jcjohnson/neural-style/issues/7

Oh, and since you're basically using this commercially you might also want to take a look at the license of the VGG model: https://gist.github.com/ksimonyan/3785162f95cd2d5fee77#file-...


Am I using it commercially? The results are given away for free. If you want s human artist to paint it (what instapainting does) you have to upload your result as a reference image yourself.

My use would fit under the CC definition of non commercial since the neural net results are handed out for free.


It's now switched to Kyle McDonald's implementation. It's a lot slower though.


*by Justin johnson


Will there be an API for this for digital-only, high resolution pics?

Some feedback too:

1) when submitting it was a bit weird that I have to upload pics in order vs explicitly 'this is my pic', 'this is my art pic'

2) a lot of people won't have a catalog of digital art to upload as their reference pic, some samples to select from would be nice

3) after uploading my source + reference pics and clicking 'submit artwork' it takes me to a login/signup page for AMANUFACTORY.COM. That button really stands out as what-you-should-do to finish the process!

4) would be nice if instead of selecting one painting it generated lots of different styles


I hacked this in one day using some UI for existing stuff, so the UI will be a bit weird.

After uploading your source + reference, you don't have to do anything. The Submit button at the bottom is UI for other people to submit a resulting image. I'll probably remove it since it's for the other stuff.

As far as API, the process is fairly expensive (uses EC2 GPU instance), so any heavy application type usage would probably be better served by setting up your own instance. We use implementations found on github (credit at the bottom of the page).


Funny, I hacked something like this together too recently, although it did take me a bit longer. Nice job!


Thanks for the info.



Quick off the mark eh? When was this idea published by those researchers, two days ago?

No wonder universities are getting super serious about IP protection/licensing/commercialisation


arxiv says it was submitted 26 Aug, so about a week ago. Near the end of the paper they explain how to implement it, so I guess it was a matter of reading the paper and sitting with caffe to toy/hack an implementation -- most of the heavy work was "already done".

Edited some typos.


> No wonder universities are getting super serious about IP protection/licensing/commercialisation

oh god, the knowledge might be free with only attribution and no revenue! the horror!


Eh how do you incentivise the creation of this knowledge given that it costs money to do?


The kids who know what's up will know where to do the research. That should be incentive enough. Considering you can't stop another organization from building on your research, the "competition" is a little hard to explain.

A university should be run like a school, not a business.


You should look up the financial statements for one of your local colleges and see if it can't be run like a business. Just one of them costs a lot of money to keep going..

That's why they have to protect IP and commercialise it. Because if they don't, the colleges can't keep teaching as it all costs so much.


I'm sure they're trying to patent it. (eg DeepMind has been patenting everything they can, even stuff like dropout and deep q-learning!) Which raises an interesting question: Instapainting is a commercial service and this is targeted at customers:

> After your artwork is generated, you are free to use it for any purpose. Upload it to Instapainting.com as an order and an artist will physically hand-paint it!

So how does that interact with the patent application? Are they liable for damages once the patents go through or do they just have to take it down?


Well that line is more like an advertisement for Instapainting. Instapainting doesn't run on this technology.

That being said, a human artist can do a much better job than these algorithms!


Instapainting doesn't run on this technology.

Aren't most convnets trained using dropout these days?

Edit: it seems the implementation that this site uses does as well:

https://github.com/kaishengtai/neuralart/blob/master/models/...

Patents need to go :(.


Wow, I need more coffee, I read your sentence as:

Aren't most convicts trained as dropouts these days?

Thought he was contracting out prisoners to paint for him...


Seriously quick. I thought I'd take a stab at it and asked what the license was for the reference implementation. Haven't even check the answer yet :-/


The implementation that's currently running is MIT licensed.


I'd be really surprised if someone isn't already hard at work on an iOS app that uses this tech. The scalability is of course a problem, but I'd pay 50 cents per photo (assuming I got a small preview of it first) or something of that order to support the costs.


There's no public list, so if you ran an image feel free to share here.


I accidentally submitted something four different times, since I thought the form wasn't loading. Does it just start as soon as you post two images, without a submit button?


I wonder what it looks like if you upload something cubist or otherwise abstract.




The art style image should be the second image. So it tried to do the artwork in the style of the photo instead.


that was quick, the paper just came out


What paper are you refering to?

edit: Leon A. Gatys, Alexander S. Ecker, Matthias Bethge, A Neural Algorithm of Artistic Style


A Neural Algorithm of Artistic Style by Leon A. Gatys, Alexander S. Ecker, Matthias Bethge

http://arxiv.org/abs/1508.06576


can we get a bigger version of the resulting image? The output is very small.


Just upgraded the implementation used. I'll see if I can increase the size without running into memory issues.

EDIT: Increased to 768px max size.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: