Hacker News Reader

Comment by ghostcluster | Hacker News Reader

Is it just me or are these "deep fakes" just not even very convincing?

I felt the same way about all the Nicholas Cage ones that were viral recently. It's so obviously a very shallow 'effect' and can't compensate for very basic differences in bone structure, hair, etc.

View on HN

Are you talking about the pornography being unconvincing or the GIF from this article?

Fake: http://svencharleer.com/blog/wp-content/uploads/2018/02/anne...

Original: http://svencharleer.com/blog/wp-content/uploads/2018/02/elke...

I could see the fake being easy to detect at high resolution, but personally I have a 0% chance of catching this gif as fake.

You have the Fake and Original mixed up.

elke-1.gif is the fake anne-1.gif is the original

You have labeled the fake and the original the wrong way around.

They'll get more convincing. Then we'll actually get better at recognising subtle differences. Then they'll get even more convincing. Someone will develop algorithms to identify those fakes. Then they'll be further developed and improved to the point that they're impossible to distinguish at every level.

> Someone will develop algorithms to identify those fakes.

One twist: don't adversarial generative networks already work by including a "fake detector" and beating it? This cat-mouse game might last, like, all of a week.

> This cat-mouse game might last, like, all of a week

I'd give it two at least.

Most of the submitted deepfakeson reddit are the first couple experiments of inexperienced amateurs trying out the technology. The worse examples are not representative of how the technology works, they usually indicate a flaw in how that particular person build their data; and the best examples are almost undistinguishable.

The technology works now, it's just at the stage where it's simple enough for a random person to build something but still tricky enough so that not everyone can build something good; it works, but not necessarily right out of the box.

This is just the beginning though. The iPhone 1. Imagine the iPhone X version of this. It's open source now; it was made by one person. Now imagine it after 10,000 people have worked on that code. This will become so streamlined and efficient you won't be able to tell at all. Some of the good ones are already extremely convincing, you just have to give it appropriate training data, choose a good model to swap with, etc.

Hard to get excited about a technique that doesn't really work well with the handwavey excuse of "it may actually work later"

Regardless of whether this is a good or bad technology to have, I can all but guarantee it will only improve with time.

Something to keep in mind is that part of the problem with these fakes is the shallowness of the training data. If rather than, say, 30k images being used - 3 million were instead - that would cause a great change in the quality.

But - that would also take a ton more processing; instead of a clip taking several hours to create, you might be looking at it taking a few days. That, or you'd have to throw a ton of money into better hardware to handle the extra load.

I've seen a number of these deepfakes, both SFW and NSFW, a lot of the unconvincing ones are done by people that haven't specified a long enough training time (12+ hours normally IIRC) or produced enough res image data sets / were testing the data for first time.

You have to find a target actor/actress that looks similar to what you are deepfaking (e.g. family member)

Kind of similar to the CG work for Blade Runner 2049?

Like everything in AI, it's probably just a matter of time now.

Some of them, unfortunately, are pretty damn convincing.

Reply on news.ycombinator.com