Nvidia EditGAN: Image Editing with Full Control from Sketches | Hacker News Reader

Nvidia EditGAN: Image Editing with Full Control from Sketches

louisbouchard.ai

77 pointsjonbaer11 comments

I wouldn't call this blogspam per se, but the original page is here and probably a better link:

https://nv-tlabs.github.io/editGAN/

If you thought filters were warping the mass perception of beauty, wait until the social media platforms have these "smarter" filters that will allow more fine-tuned parametered control over pictures.

You can control your own pictures with character editor akin to an RPG.

AOL had the right idea with avatars, back in the day - nobody wants to share what they look like, but instead, what they _want_ to look like.

On the internet, nobody knows your an ugly dog, except your new tech overlords.

Interesting point, but I should point out that you've already been able to do this for several years now, for example with Meitu

Honestly the examples shown are underwhelming.

But the idea and technology are promising. It's like Photoshop but slightly easier and more realistic (I bet the neural network accounts for lighting and other minor details). There's a solid usecase for this kind of tool.

The examples show them realistically editing almost any feature of a face, cat, car, painting... Human, animal, object, art.

What would be a better example or usecase?

It's quite depressing that in 10-15 years photographs and video won't really be admissible as court evidence in most cases. At least they won't be any more credible than a witness statement. Photography and especially video has been a great equalizer when it comes to accessing justice.

We'll be able to use cryptographic signatures to prove who said what. Once people become less able to mindlessly watch video from whoever they like to believe, and have to actually start paying attention to such signatures and the reputations of the entities signing .. uh, not sure how to finish that sentence but it might be good

I’ve thought about this too. Each image in a chain of altered images could be cryptographically signed / verified, that makes sense. The hardest part seems like verifying the original “straight out of camera” video/photo, maybe hardware manufacturers implement a way to prove the output is original?

Even if we can't get workable signatures from cameras, we can create workable signatures by reporters on videos they took. A reporter's reputation will become valuable. Currently (many) people believe video even when it's posted by shameless, obvious liars, because hey, the video can't lie, right? Once it can, maybe they'll actually care more about the error record of the news outlet they consume from.

It's still not a great situation, because it means when a casual passerby takes an important video for the first time, it will have trouble getting any traction. But maybe it's at least a silver lining.

Can you use the output of this in a commercial product? Are any of these machine learning image manipulation programs permissive like that?

The memes are going to be epic.

Reply on news.ycombinator.com