Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

junyanz.github.io

Post reply on HN
Print view View original

31–40 of 146 posts

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#31

Post by cr0sh »

jimmies wrote: post #23

While this is undoubtedly very impressive, I think it's just another logical step to what we've been seeing so far: - In the past, you needed to have a pianist at home to perform you a song, with the music box and then the phonograph you don't need to hire anyone anymore. It's probably not as good as a live performance (maybe?), but it's good enough for many people, and much much cheaper, faster, and available. - You…

The next step would be to have it actually turned into an actual canvas rendering, with the texture and such of actual paint. Could probably be done using 3d printer technology, inkjet, and/or robotics in some manner...

...in fact, I know you can already get photos printed to canvas - but taking it to the next stage of texture would be amazing - right now, I think the best you can get is to have a trained person "highlight" areas of the canvas with paint. From what I understand, there's a whole "village" or small city in China that specializes in custom painted images (which a lot of online places use); I would be surprised if there isn't an effort to automate this work.

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#32

Post by rememberlenny »

king_magic wrote: post #7

This is really incredible. Maybe a neat idea: artist draws/paints frames for animation, frames are converted into semi-photorealistic images through this software, and assembled into a movie.

I am a painter and did something similar. I took two images, one of nature and another of street art. Using style transfer, I combined a series of images. After identifying a few results that I liked, I used actual paint to make large canvases inspired by the digitally produced images.

Link: https://twitter.com/rememberlenny/status/825026441603592193

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#35

Post by cr0sh »

eanzenberg wrote: post #29
thriftwy wrote: post #6

The next step is "turning pencil drawings into photos" and using it to fabricate evidence on grand scale. Why bother catching politicians doing something when you can just draw them in? Will wreak havoc on societies with weak politics/reporting culture.

So... photoshop?

I think beyond photoshop - more like the ability to seamlessly insert or "imagine" situations that never occurred in real life. For instance, imagine if you could feed in a ton of images of person A - then a ton of images of porn - then have the system imagine a porn scene.

...or a drug-usage scene.

...or any other thing you could think up. It's even possible this could be pushed into a generated movie. You could even work the "voices" in (have the person say virtually anything you want, nearly perfectly, synched perfectly with the facial expressions/movements/body language).

There's a lot of scary, yet interesting possibilities here!

Hmm - think of the possibilities for CGI effects? Rather than having to build a 3D model, just have the software imagine the scene (I know - much easier said than done - but I think the possibility is there, now).

Another idea: Could this kind of software take a movie or still image, and "imagine" a three-d model from it? I don't see why not...

...and yes, I understand this software as being a particular form of a neural network, and that for any of this it takes massive compute power in the form of multiple-GPU time and a ton of memory - and likely isn't yet in the realm of doing it yourself at home except in simpler examples - which might still take days or weeks to train, and a good amount of time to render; I just want to indicate that I have a certain minor level of understanding of what this software is about and that I don't believe for a second that it is a basic "photoshop" filtering system.

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#36

Post by keehun »

Terribledactyl wrote: post #19

My favorite of the results tended to be [anything] -> Ukiyo-e/Cezanne. I think because these are easier problems, lots of detail to less. The transfiguration and painting -> photo have me firmly in the uncanny valley, but I suspect this harder problem will be solved given more training.

I also really liked the Ukiyo-e outputs. I wish they would release full-resolutions at some point!

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#40

Post by xigency »

jordache wrote: post #27

TIL - remap green pixels to grey/white and you go from summer to a contrived notion of winter. This done easily in Photoshop. - select green pixels, smooth it a bit, then paint white over it.. - apply a blue cast on it https://s17.postimg.org/q68dz04sf/test.jpg

Okay, now do the Zebra one.
Post reply on HN