I've been wondering what is going to happen when bots start using these images as profile pictures?
Previously it wasn't trivial to do a GAN image generator, now as this site shows it's, if not trivial, also not particularly hard.
Previously it wasn't trivial to do a GAN image generator, now as this site shows it's, if not trivial, also not particularly hard.
... but all, or almost all, of these would be irrelevant at a profile pic size. At that size, assuming these aren't just recapitulations of the training data (and I assume they aren't) this technique appears to be 99%+ successful.
Also, look at the non-face stuff. Some backgrounds are just "incredibly blurred vaguely landscapy stuff", which is plenty realistic, but I've seen the algorithm attempt wood grain, which went poorly. I've seem some bizarre patchwork backgrounds, and one picture had a person cut off to the right like a single photo trimmed from a family photo, and the cut-off person was some sort of SCP-monstrosity mercifully cut off by the edge of the photo. Still, the success is impressive. The failures are definitely going from "in your face" to "easy to ignore/miss".
Fix up the training data a bit and this'd be a profile pic machine.
Every single picture, if you really look at it, is disturbing for reasons you can't pick out. It's definitely hitting the uncanny valley. It's juuuust human enough to blend, but not human enough to avoid the creeping feeling of dread.
But that being said, if I'm cruising forums and see this in a thumbnail size, I'm not going to be able to pick out that it's not a real person.
I'd say you're correct that it is still in it, but it is clearly climbing up the other side now. We're past the minima of verisimilitude now.
https://thenewstack.io/deep-learning-ai-generates-realistic-...
The biggest problem is transferring faces to existing photos. It was hard to do manually. Now it's much easier. Also, people are generally trained to ignore various artifacts by CGI-ridden movies and compression algorithms. So much of our notion of how the world looks comes from digital imagery, it's kind of scary.
I think we need to change the threshold of quality for an image/video to constitute "proof" of any kind. You can hide most of the weird artifacts by scaling things down or passing them through heavy compression.
The generator is like 150MB. The forward pass is <0.1s. Hypothetically you should be able to generate on a decent GPU like a 1080ti with 11GB VRAM at full utilization <730 images per second. Use a few GPUs and you're at thousands per second.