Amazing. And now one only needs a random article generator, which describes marriages, breakups, mishaps, accidents, and plain random appearances of these celebrities. Tada, random fake celebrity news, which is probably even better at wasting people's time than the real thing. This is going to happen really soon, because text generation is much easier than what the Nvidia guys did.
Just came here to say that it's time to start a gossip magazine about completely imaginary people. You beat me to it.
However I disagree about text generation, which seems to me to require a perfect equivalent of human intelligence to be at least acceptable. (I'd call it AGI-complete, or maybe HI-complete).
I think these people being "real" is a secondary concern. Most readers of gossip magazines only see the characters from them only in the same magazines, or TV. Soap opears about completely imaginary characters faced a lot of success. I don't see why algorithmically generated soap operas won't have a chance, especially when adorned with "photos".
Of course I meant it as a joke. But I think the problem- or at least the point of the joke- isn't much in the fact that you haven't seen these people before. The point is rather that "gossip" is a sparse amount of information about some reality you don't have direct access to, but that exists and is consistent. So with normal gossips, it makes sense for the brain to gather all this info and try to infer a picture of the person behind it. However, if there is no real person behind the gossip the entire exercise is perfectly illusory.
Repeat a lie enough and it becomes the truth; they just have to make those fake celebs appear in TV now and then to turn them into cash cows.
I'm 100% sure in due time we'll see Elvis singing stuff that Elvis has never sung in his life; producing fake celebs would be not that different: much less initial profits but also much less royalties to pay.
No threat of lawyers, no recourse, no boring luls, no lack of luxery/perversion and punnishment patterns, which allows for envy to be disguised as moral outrage - synthetic gossip is better then the real thing.
We allready have revived and virtual stars on the stage.
Well, create a few templates like some online media currently use and let bots fill the info in. The content should not change much from event to event.
I think that celebrity gossip articles would be prove to be as ritualized and standardized as sports and stock market articles; places where algorithmic article generation has worked well. You'd probably only have to hire human writers to do longform features and interviews, and to create an underlying model to generate the activity of these fake celebrities that you would be reporting on.
The thumbstamp version of creating fake celeb faces was one of the projects for the Deep Learning Udacity nanodegree. (These get changed over time, so it may be different now.)
The dataset of 200,000 celeb photos with the face nicely centered at a known location is a nontrivial part of making the exercise feasible.
I trained on windows with a 6Gb 1060, and went off script from the DCGAN paper, by using upscaling rather than transpose convolutions. Once all the fiddling details are set correctly, the results are quite amazing. It didn’t even require a complete single pass over that dataset.
Really interesting article seeing how I'm finishing up my own DCGAN project!
Generative models like GANs are fascinating, but very temperamental to train. Some of the findings in this paper mirror my own observations - increasing the complexity of the GAN adds a lot of instability. My solution was to keep things as simple as possible. I spent a lot of effort trying to increase the size of the network to get better results, but in the end my smallest implementation worked the best.
This bit is interesting: "Without progressive growing, all
layers of the generator and discriminator are tasked with simultaneously finding succinct intermediate representations for both the large-scale variation and the small-scale detail. With progressive
growing, however, the existing low-resolution layers are likely to have already converged early on, so the networks are only tasked with refining the representations by increasingly smaller-scale effects as new layers are introduce"
I read the paper, but I did not understand a thing. What is the path to follow (for example, books or papers to read) in order to at least understand what the paper is talking about? My background is in computer science.
I did the Udacity nanodegree last year and the only part I never understood was the call to the "de-convolution" operator in tensorflow. It seems that every description keeps painting the same picture of how the two networks are in competition to reach "nash equilibrium". I think they really skimp over what a "de-convolution" actually is.
Did you do the DLND? If so, they may have updated the videos, as the current explanation of conv2d_transpose seemed fairly clear to me. Or maybe it was some of the extra materials they provided that helped. I think I found a good video on YT also.
From the demonstration video [0] in the link, very interesting, but also very noticeable where the GAN attempts glasses or other "items" (see around 0:58 and 2:35, respectively). Has any research been done into GANs which aim for realistic artificial objects, instead of faces?
It is interesting really studying the two images on the github page. My first thought was - wow, amazing! But having looked at them for a bit longer they've fallen straight into the uncanny valley for me.
The female's left and right eyes are different shapes, as are her eyebrows. And the male's ears appear to be in different places on the left and right side of his head. His eyes are creepily different too.
Well, my ears are on different places, too, which is why glasses were always a bad fit before I got PRK.
I’d say that the shape of the woman’s hair feels weird, but I’m not really offput by the image of the man.
