Sometimes I wonder at the uniqueness of each individual's face. It just amazes me how everybody, no matter how numerous we have, has a different face. With the exception of twins and non-familial coincidental twins.
Practically speaking, every face has to communicate, what, 20 bits of information, to avoid common doppelgängers via the birthday paradox? Probably even more. Very interesting.
People have proposed leveraging this to generate unique avatars, since you can just pick some random parameters for the arrangement of a stick figure face and it will probably look unique.
Openface (probably state-of-the-art currently? If I haven't missed anything) uses a 128-dimensional unit hypersphere to represent a face identity. They have likely optimized this, so it's the dimensionality that works well in practice, not too large, not too small. It's perhaps a bit bigger than it has to be, but it cannot be too big either because that would very likely lead to overfitting.
So that's 128*sizeof(float) = 512bytes = 4096 bits of information.
Kinda like an RSA private key :).
From an artist's point of view, lots of small details and imperfections and things like that. Many bodies have basically the same proportions - your hand is around the same length as the distance between your chin and eyebrows, eyes are an eyeball length apart, nose stops approximately halfway between the corner of the eyes and the chin.
But you probably have one eye slightly higher than the other, one eyebrow that is shaped different, or your nose might turn up ever-so-slightly. Your cheeks might show more "chub" when you smile compared to the next person, even though the weight is similar. I think this is why painted or drawn portraits sometimes don't quite capture someone's face, even though it seems horribly close and is obviously a picture of the subject - and weirdly, why simple caricatures seem to do such a good job. The folks that do them have trained themselves to notice and highlight these sorts of differences.
As a sidenote, I still find it amazing that the human brain is so quick to pick up such things, especially considering it takes training and practice to reproduce on paper.
I love the dream-like quality these faces have, even the "ugly" ones. They look like something out of Baldur's Gate and somehow make my imagination spin. Do you know how the effect is called ? I've experienced the same thing when looking at screenshots from Elemental: War of Magic (but it was hit and miss) and world map in Shroud of Avatar (the new project of Richard Garriott, the designer of Ultima serries... I think it's finished now).
Note I don't intend to promote any of the above games except BG. I just have a soft spot for this style of graphics.
Heh, my first idea too was "these portraits would look awesome in RPG". Furthermore, automatically generated art could help indie game creators, as such games often suffer from generic looks and hiring real artists is usually over the budget.
My wife is a portrait artist, and when she usually does a commissioned portrait for someone, I am astounded by the amount of time that she spends purely working on the eyes. (ex. [1],[2])
Looking at these in comparison to her work, it is definitely true that the eyes are the windows into the soul, and define a LOT about a person's face. (Or even an animal - she does pet portraits too).
Even in looking at her very early work in art school - I can forgive the odd crooked nose or askew mouth as long as the eyes tell the story.
Looking at these AI generated portraits, I find the facial features really interesting, but it is the artificial, dead quality of the eyes that make me somewhat uncomfortable with them, and induce a slight feeling of horror, as exceptional as the artwork is.
The most interesting thing about the procedurally-generated portraits (and earlier work on other generated art [1]) is that the the results completely squick out some people while other people don't seem to notice what the problem is and think we'll be drowning in 4k procedurally-generated Ghibli movies on Netflix in a few decades from now (see [2]). Clearly the two groups are looking at the pictures in two completely different ways.
(Personally, I tend to fall into the 'squicked out' category. I don't like the way objects in [1] melt into each other, like the artist is a lobotomy patient who can no longer separate objects, or like the picture is depicting a hideous Lovecraftian fractal dimension.
The faces are actually better because the author iteratively upscales something that looks like a face at every stage, instead of just running one step to generate an alien blob of too many eyes and noses that happens to pass the neural network. Besides dead eyes, they tend to suffer from the problem of the two halves of the face trying to make completely different expressions.)
I find these images fascinating, specially the parts were the textures blend with the faces. It feels like we are watching the world through another mind, so to speak.
15 comments
[ 3.9 ms ] story [ 49.2 ms ] threadPeople have proposed leveraging this to generate unique avatars, since you can just pick some random parameters for the arrangement of a stick figure face and it will probably look unique.
But you probably have one eye slightly higher than the other, one eyebrow that is shaped different, or your nose might turn up ever-so-slightly. Your cheeks might show more "chub" when you smile compared to the next person, even though the weight is similar. I think this is why painted or drawn portraits sometimes don't quite capture someone's face, even though it seems horribly close and is obviously a picture of the subject - and weirdly, why simple caricatures seem to do such a good job. The folks that do them have trained themselves to notice and highlight these sorts of differences.
As a sidenote, I still find it amazing that the human brain is so quick to pick up such things, especially considering it takes training and practice to reproduce on paper.
Note I don't intend to promote any of the above games except BG. I just have a soft spot for this style of graphics.
Looking at these in comparison to her work, it is definitely true that the eyes are the windows into the soul, and define a LOT about a person's face. (Or even an animal - she does pet portraits too).
Even in looking at her very early work in art school - I can forgive the odd crooked nose or askew mouth as long as the eyes tell the story.
Looking at these AI generated portraits, I find the facial features really interesting, but it is the artificial, dead quality of the eyes that make me somewhat uncomfortable with them, and induce a slight feeling of horror, as exceptional as the artwork is.
[1] - https://www.instagram.com/p/BRhU7fKjdhS/
[2] - https://www.instagram.com/p/BRZormMDyB3/
(Personally, I tend to fall into the 'squicked out' category. I don't like the way objects in [1] melt into each other, like the artist is a lobotomy patient who can no longer separate objects, or like the picture is depicting a hideous Lovecraftian fractal dimension.
The faces are actually better because the author iteratively upscales something that looks like a face at every stage, instead of just running one step to generate an alien blob of too many eyes and noses that happens to pass the neural network. Besides dead eyes, they tend to suffer from the problem of the two halves of the face trying to make completely different expressions.)
[1] - https://research.googleblog.com/2015/06/inceptionism-going-d...
[2] - https://news.ycombinator.com/item?id=14510453