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

Creating a deepfake took two weeks and cost $552

arstechnica.com

Post reply on HN
Print view View original

61–70 of 119 posts

Re: Creating a deepfake took two weeks and cost $552

#61

Post by everdrive »

simias wrote: post #4

A few weeks ago my girlfriend made me discover Instagram filters. Obviously they're not in "deep fake" territory but I was still massively impressed by the quality of the face recognition and tracking coupled with some rather advanced 3D reconstructions on top. 15 years ago that would've been science fiction, today it's a free gimmicky feature in an app running on commodity hardware. There are also apps that will smo…

I'm old enough to remember when people still attempted to claim that fashion magazines were forcing airbrushing on women. Clearly, there's a large amount of market interest for this sort of capability.

Re: Creating a deepfake took two weeks and cost $552

#62

Post by claytonius »

tralarpa wrote: post #56
gdubs wrote: post #50

The central technique used in DeepFakes is fascinating. Initially I assumed they were using Generative Adversarial Networks (GANs). But — and someone can correct if I’m wrong — Deep Fakes use two Autoencoders. So essentially you’re training a network how to compress an image down to a very tiny representation, and then uncompress it as accurately to the original as possible. You train two of these: one for the origin…

How do you guarantee that the two trained autoencoders speak the "same language", i.e. the code learned by the autoencoder of the original face gives a reasonable output when uncompressed by the target autoencoder?

Something like a CycleGAN and/ or weight sharing would do the trick

Re: Creating a deepfake took two weeks and cost $552

#63

Post by jldugger »

binarybits wrote: post #26
thathndude wrote: post #3

I mean. No. I understand the point of the article is “look how much I did with so little.” And it’s really good in terms of how much was done. But the final deliverable isn’t at all convincing. Clickbait title.

Man, you're reading way too much into my headline. The headline describes the article accurately: I created a deepfake. It took me two weeks and I spent $552. The headline doesn't say anything about whether it was convincing or not.

> The headline doesn't say anything about whether it was convincing or not.

I guess I always assumed the 'deep' in deepfake meant it was convincing, but upon further consideration it's probably just a nod to DNNs. Well played sir.

Re: Creating a deepfake took two weeks and cost $552

#64

Post by orange3xchicken »

tralarpa wrote: post #56
gdubs wrote: post #50

The central technique used in DeepFakes is fascinating. Initially I assumed they were using Generative Adversarial Networks (GANs). But — and someone can correct if I’m wrong — Deep Fakes use two Autoencoders. So essentially you’re training a network how to compress an image down to a very tiny representation, and then uncompress it as accurately to the original as possible. You train two of these: one for the origin…

How do you guarantee that the two trained autoencoders speak the "same language", i.e. the code learned by the autoencoder of the original face gives a reasonable output when uncompressed by the target autoencoder?

This is a good question. I'm not sure which paper the parent comment is referencing - and in general, DeepFakes is a very general term for a variety of techniques including GANs. The problem you describe is called the alignment problem and it's studied in several areas of applied ml: translation, multi-modal learning, etc.

For example, the canonical way to train deep machine translation systems is to make use of large parallel corpora - large set of translation pairs. In the case where this data may not exist, for structurally similar languages, one approach is to learn unsupervised word embeddings independently (i.e. using word2vec) which only require monolingual data and to find a transformation that aligns the two spaces. In many cases, neural networks still struggle with this task. However, some work in the last few years has demonstrated that a few seed word-translations is all you need to apply simple algorithms that can learn a linear or simple nonlinear transformation that work pretty well!

[1] Here is a method that proposes locally linear transformation maps for alignment: http://nakashole.com/papers/2018-emnlp-norm.pdf

Re: Creating a deepfake took two weeks and cost $552

#65

Post by gnicholas »

xedeon wrote: post #9

That's a terrible Deepfake to be honest. This is a much better example. https://www.youtube.com/watch?v=bPhUhypV27w

Agreed. The eyes in this video look weirdly bloodshot and unnatural. Even before it started playing I thought it looked bizarre.

Re: Creating a deepfake took two weeks and cost $552

#66

Post by seriousaccount »

panarky wrote: post #58
trianglem wrote: post #37

Earlier quoted context omitted.

I feel like that ship sailed a while ago with Photoshop being around. I think the problem now is you can do the same with videos and the technology is more commoditized so you don’t have to be very knowledgeable to do these things.

People predicted that Photoshop was the end of truth, that journalism could no longer be trusted, that photographic evidence at criminal trials would be inadmissible. None of that actually happened because people learned to be skeptical, and to care about who took the photo, the chain of custody, and the controls to prevent digital alteration. Maybe we need something similar but more rigorous for video. The camera it…

I don't agree with your assessment. People didn't learn to be skeptical most of the time. We've seen what fake news does to people's opinions; debunking a video as fake after it was viewed by a large group of people is enough to do serious damage to people's viewpoints.

Re: Creating a deepfake took two weeks and cost $552

#67

Post by creaghpatr »

Pandabob wrote: post #22
simias wrote: post #4

A few weeks ago my girlfriend made me discover Instagram filters. Obviously they're not in "deep fake" territory but I was still massively impressed by the quality of the face recognition and tracking coupled with some rather advanced 3D reconstructions on top. 15 years ago that would've been science fiction, today it's a free gimmicky feature in an app running on commodity hardware. There are also apps that will smo…

It's my understanding (and experience) that Snapchat filters are even more impressive and that Snap CEO Evan Spiegel has pretty much bet the company on AR [1]. [1]: https://www.theverge.com/2019/10/22/20927521/snap-growth-ear...

The snapchat filters have been pretty mindblowing. I don't use them often but it's wild how far they've come in a few months.

Re: Creating a deepfake took two weeks and cost $552

#68

Post by ghaff »

gnulinux wrote: post #16
darkcha0s wrote: post #7

Earlier quoted context omitted.

Made by a dude who has next to no experience in the field. Now imagine what a guy who knows what he's doing can do with a bit more money. That's what the allusion here is; the fact that a Reporter with 500$ even comes close should be an alarm signal to you.

I'm under the impression that these things are easier to come infinitesimally close to, but very hard to get right: i.e. the old uncanny valley problem. A reporter with $500 can get very close, but a team of PhD with thousands of dollars and a lot of time may never be able to produce something that'll convince the public. I'm NOT claiming that this is the case, I'm just saying we don't have a reason to believe why th…

I'd really hesitate to bet on never in this case--even if never were scoped to mean not within 20 years.

There really are two separate questions though:

- Can experts with a big budget create videos that even equally expert analysis can't tell from the real thing? (Maybe yes, maybe no. But it's hard.

- With modest skill and budget, are people increasingly able to create fake videos that stand up a cursory glance and can fool people as a result? (Absolutely--or at least we're getting close.)

Re: Creating a deepfake took two weeks and cost $552

#69

Post by trianglem »

yboris wrote: post #33
darkcha0s wrote: post #7

Earlier quoted context omitted.

Made by a dude who has next to no experience in the field. Now imagine what a guy who knows what he's doing can do with a bit more money. That's what the allusion here is; the fact that a Reporter with 500$ even comes close should be an alarm signal to you.

This is also an impressive DeepFake: https://www.youtube.com/watch?v=Wm3squcz7Aw Voice actor reads a poem in voices of famous people; his face gets swapped with the actor's face; the combination of voice and face makes it very believable.

This one is fantastic. It seems like just the most characteristic traits are carried over and the result is seem less.

Re: Creating a deepfake took two weeks and cost $552

#70

Post by bryanrasmussen »

goatlover wrote: post #29
bryanrasmussen wrote: post #2

Well Zuckerberg and Data do look like they were separated at birth, which I guess would make Zuckerberg Lore - which might also explain a few things.

So FB is the crystalline entity or the rogue Borg?

Facebook does seem to have a strategy of assimilating every successful social platform.
Post reply on HN