Around twenty-five years ago my elder son passed his GCSE Art with a project, rather reluctantly endorsed by his art teacher, using ray tracing to produce drawings which he had generated on my computer, I think at the time still an Amstrad PC1512. I think we had to leave it running overnight for the best of them, and could only produce hard-copy by photographing the screen as we only had a black and white printer, but my memory of them is faint.
They weren’t particularly good drawings, though better than some entered for GCSE Art, and were certainly in no way photo-realistic. But both computer hardware and techniques have made great bounds since then, and the latest faces generated using AI and shown on PetaPixel in These Portraits Were Made by AI: None of These People Exist are entirely convincing.
The were produced by NVIDIA researchers Tero Karras, Samuli Laine and Timo Aila using generative adversarial networks (GAN), about which even they write “Yet the generators continue to operate as blackboxes, and despite recent efforts, the understanding of various aspects of the image synthesis process, e.g., the origin of stochastic features, is still lacking.” Having briefly scanned their publication, which contains the images shown on PetaPixel, my understanding is still definitely lacking, and, unless you are the kind of person who crunches tricky equations before breakfast it is unlikely to add much to your comprehension either.
I can’t even get this blog to reproduce the equations properly, but here’s one I just found:
lZ = E h 1 2 d G(slerp(z1, z2; t)), G(slerp(z1, z2; t + )) i
where z1, z2 ∼ P(z), t ∼ U(0, 1), G is the generator (i.e. g ◦f for style-based networks), and d(·, ·) evaluates the perceptual distance between the resulting images. Here slerp denotes the spherical interpolation operation [49] that is the most appropriate way of interpolating in our normalized input latent space…
So I guess that makes it all clear?
What I can see is the potential that this development has for fake news and for advertising images (and as another group of images from the paper illustrated entirely filling the few gaps in Facebook not already occupied by cat pictures.)
Doubtless it won’t be long before programs based on this a other similar research are common on our desktops (and even on our phones) and as well as producing non-people will be churning out images of real people doing things they never did in places they never visited.
I’m unsure too, about the copyright issues involved around these images, which rely on multiple real photographs for their generation, though I suspect those who run the software will claim the copyright.
Nor is it easy to predict the effect it will have on photographers, though it has the potential to replace much of the stock photography market, something that would not greatly worry me, though I think may greatly reduce employment in the area.
It may even increase the value of the ‘real’ photograph, an image whose integrity is vouched for by the credit line of the photographer – so long as we retain our integrity and our photographs have something to say.