News for hackers. Stuff that matters.
Re:Imaginary worlds dreamed by BigGAN by bufferoverflow on Saturday November 17, 2018 @01:14AM (#18474261)
RIP stock photography. And possibly videography too.
Re:Imaginary worlds dreamed by BigGAN by mkagenius on Saturday November 17, 2018 @01:29AM (#18474291)
Very doubtful, the results just seem to be combination of existing background and existing artifacts (nose, ears, furs, etc) .

The different poses you would get naturally is much more complicated and diverse than these networks seem to be generating. All in all these networks aren't generating anything new.

Re:Imaginary worlds dreamed by BigGAN by visarga on Saturday November 17, 2018 @02:00AM (#18474371)
The images you saw in the article were hand picked to be weird. Most of them are photorealistic and are not copies of any training photo. If you think it's so easy, paint a photorealistic image in Photoshop by hand (no direct copying, but you can look at other images all you want) and let's see how it compares.

Cropped a few more BigGAN samples for reference: https://imgur.com/a/GIQSo4I

Or how about ProGAN for faces? https://www.youtube.com/watch?v=36lE9tV9vm0 Does that look like a simple copy? It's a 'walk in the latent space' of faces.

Also, take a look at this tool (TL-GAN) https://www.youtube.com/watch?v=O1by05eX424

That comes with an online demo: https://www.kaggle.com/summitkwan/tl-gan-demo

Re:Imaginary worlds dreamed by BigGAN by mkagenius on Saturday November 17, 2018 @03:37AM (#18474626)
> photorealistic image in Photoshop by hand

If you are competing with a human, if not me I am sure a professional can easily better the best of best (and not just 512x512 pixels but 100 time more pixels) by this network. It has to catch up with humans by a long shot.

> Cropped a few more BigGAN samples for reference: https://imgur.com/a/GIQSo4I

I don't know but to me all of the images you have in the imgur link seem non realistic. Head or torso or legs is messed up for (all?)most of them.

> how about ProGAN for faces? https://www.youtube.com/watch?v=36lE9tV9vm0 Does that look like a simple copy? It's a 'walk in the latent space' of faces.

No, I am not saying its a simple copy, the copy is great. The fact is that its a copy. An imperfect one.

> Also, take a look at this tool (TL-GAN) https://www.youtube.com/watch?v=O1by05eX424

Sure, human face comes in all shapes and sizes, so its tricky to judge the effectiveness since every modification is a possible real person.

Re:Imaginary worlds dreamed by BigGAN by visarga on Sunday November 18, 2018 @11:41AM (#18481181)
> Sure, human face comes in all shapes and sizes, so its tricky to judge the effectiveness since every modification is a possible real person.

What I wanted to show you is that you can walk through the 'latent space' of faces and generate intermediary images between any two images, thus they are not simple copies.

Re:Imaginary worlds dreamed by BigGAN by anonytrary on Saturday November 17, 2018 @02:07AM (#18474396)
If you train a NN with images, wouldn't you would need permission to use the training images in the first place?
Re:Imaginary worlds dreamed by BigGAN by sgt101 on Saturday November 17, 2018 @04:38AM (#18474778)
I think that that will depend on the license and the courts. A NN can be seen as a storage system - these images are retrieved from the compressed and transformed data stored in the NN; a court might rule that that is essentially the same as any other storage system. Or not.