Would be interesting to test against real photos that are similarly zoomed in than the photos from TPDNE. Also, having the eyes at the same position in the image will be a great test, as the eyes are always at the exact same position for TPDNE images. And if their AI trained on that aspect, it might be a "eyes are at the same position"-detector
Interesting that it can't detect the face-swap/deepfake or non-human subject. I suppose you could get that result if the network was simply trained to sort in binary: "real" or "fake", rather than into "real", "fake", and "not human". That, of course, would make the problem significantly more challenging and necessitate more forethought into how to do it ethically. Thanks for braving the invasive permissions and sharing this! Also just fyi, there is an editing bug around 5:04.
Overall pretty neat, but in your testing most of the images were taken at different angles than most AI generated photos I've seen. Yours in particular was taken from farther out, the following one was a side shot in really quite unique lighting, (I like it at least) and the fourth one was also taken from farther out additionally with a B/W filter. The third one is the only one that I could reasonably see being a result of an AI generation from a cinematography perspective but all the others are noticeably shot different. If the AI DID factor in things like cinematography(either directly or indirectly) you could argue that that is a valid technique since, well, yeah if it's shot differently it's probably not fake, but an over reliance on that sort of shot cinematography could lead to false positives when people happen to just take a picture of themselves in a similar way to that of the AI generated faces. (which, lets be real, it's a pretty generic style, most people without a large following are probably quite close to that sort of style) Not saying that the AI DOES have that flaw, or even that if it did it would be critically important, but it's something my overly cynical mind noticed when you were checking against known real/fake images.
So I really want to see you throw some photoshop filtered GAN through it, say like, grab the GAN image using an image capture (not downloaded, to avoid any hidden metadata), then run it through say, a 2px static and blur filter. I have a feeling it won't be able to detect these?
Even if it gets better at deep fakes and AI generated photos a much simpler, and probably more useful tool, would just be a plug in that searched for original sources. I think, at least right now, most fake accounts are probably still using 'real' pictures stolen from other people's accounts. /Anyone else half expecting Jordan's entire channel is a deep fake and that she's actually a robot?
Would be interesting to test against real photos that are similarly zoomed in than the photos from TPDNE. Also, having the eyes at the same position in the image will be a great test, as the eyes are always at the exact same position for TPDNE images. And if their AI trained on that aspect, it might be a "eyes are at the same position"-detector
Interesting that it can't detect the face-swap/deepfake or non-human subject. I suppose you could get that result if the network was simply trained to sort in binary: "real" or "fake", rather than into "real", "fake", and "not human". That, of course, would make the problem significantly more challenging and necessitate more forethought into how to do it ethically. Thanks for braving the invasive permissions and sharing this!
Also just fyi, there is an editing bug around 5:04.
Overall pretty neat, but in your testing most of the images were taken at different angles than most AI generated photos I've seen. Yours in particular was taken from farther out, the following one was a side shot in really quite unique lighting, (I like it at least) and the fourth one was also taken from farther out additionally with a B/W filter. The third one is the only one that I could reasonably see being a result of an AI generation from a cinematography perspective but all the others are noticeably shot different. If the AI DID factor in things like cinematography(either directly or indirectly) you could argue that that is a valid technique since, well, yeah if it's shot differently it's probably not fake, but an over reliance on that sort of shot cinematography could lead to false positives when people happen to just take a picture of themselves in a similar way to that of the AI generated faces. (which, lets be real, it's a pretty generic style, most people without a large following are probably quite close to that sort of style) Not saying that the AI DOES have that flaw, or even that if it did it would be critically important, but it's something my overly cynical mind noticed when you were checking against known real/fake images.
So I really want to see you throw some photoshop filtered GAN through it, say like, grab the GAN image using an image capture (not downloaded, to avoid any hidden metadata), then run it through say, a 2px static and blur filter.
I have a feeling it won't be able to detect these?
Yeah, another easy test would be to try cropping or doing perspective transformations to the ai faces and seeing if they still get detected
Fascinating!
Fact checking faces… the mating game has become more interesting. Thanks JH.
Please show us how macbook m2 air performance while training Ai model i thinking to buy one because they say it has 40% enhanced neural engine
how odd that i found your channel through the tom scott video and this is the first one i watch
I really like your videos because I always learn something new. Thanks 👍🏿
Will these tools also be used to train the AI generating deep fakes to make it less detectable? It could become and arms race.
Even if it gets better at deep fakes and AI generated photos a much simpler, and probably more useful tool, would just be a plug in that searched for original sources. I think, at least right now, most fake accounts are probably still using 'real' pictures stolen from other people's accounts.
/Anyone else half expecting Jordan's entire channel is a deep fake and that she's actually a robot?