3 ms·
> Can we please get real nerdy about this? So as someone who works in generative modeling I'll give my best guess as to what is done and what is happening. It
by godelski 4y ago
> Can we please get real nerdy about this?
So as someone who works in generative modeling I'll give my best guess as to what is done and what is happening. It is a guess because they don't say everything, but there are some hints. Scambier linked these two twitter threads[0][1] which can give us some insight.
> I'll explain negative prompt weights, in case you don't know. With these, instead of creating an image of the text prompt, the AI tries to make the image look as different from the prompt as possible.
What's important here is that the machine doesn't actually know what the opposite is. In fact, I would ague we don't either. What you can do is use Lp distances from a latent representation. This is where things start to make sense. If in that we find faces as a large distance away, it is also unsurprising that we find many different facial characteristics. These first images look like there is a high mixture between strong masculine features and strong feminine features. These are not things we typically see in reality and combine with our hyperactive brains for recognizing other human faces, we enter the uncanny valley.
Next I don't know if this was done on purpose or not, but there are very clear issues with scene lighting. I can totally believe that this is not on purpose because this is something generators are bad at already. So we have shadows cast along the face in unnatural ways. Upping the creepiness factors.
Now we need to look at important features for recognizing faces: eyes, mouth, and nose. You may have noticed that text to image generators are typically really bad at these. Generators are also typically bad at facial symmetry (why we're trying to get transformers in, but this still isn't working to the degree we would like). In fact, I actually find it more interesting that these are coherent given the explanation of how the latent representation was created.
So I think we have good explanations as to why this would turn creepy very fast. Especially given the hype and that the creator is leaning into it. But these are my best guesses. I can't really know without seeing what is done.
But honestly, I am super interested and would like to see these latent representations and play around with them. This could be a good thing to investigate if you are trying to determine how smooth the latent manifold is, which is extremely important if we're going to make deeper content contributions and rely less on our prompt engineering. Maybe I'll have to play with some negative prompts (if I can find the time lol).
[0] https://twitter.com/supercomposite/status/1567162288087470081 https://twitter.com/supercomposite/status/156716228808747008...
[1]https://twitter.com/sheslostheplot/status/1567370919487893504 https://twitter.com/sheslostheplot/status/156737091948789350...
- StrictDabbler 4y agoMy first reaction was that it's surprisingly close to the traditional "average face" [0] generated by overlaying photos that we've seen in so many articles. My second reaction was to look for articles that don't average out young faces and the result gets even closer. [1] The traditional simple averaging process removes most wrinkles, rosacea and blemishes because those differ between individuals but the facial proportions match Loab well. It appears the AI when negatively weighted gives the most average possible result that still matches the concept of a face while picking the most unpopular levels of skin texture and lighting. It's basically Courtney Cox at 70 after a ten-year whiskey bender. [0] https://petapixel.com/2011/02/11/average-faces-of-women-in-40-countries/ https://petapixel.com/2011/02/11/average-faces-of-women-in-4... [1] https://www.researchgate.net/figure/Average-faces-of-young-20-29-and-old-60-69-women-from-French-Caucasian-Chinese_fig1_318677248 https://www.researchgate.net/figure/Average-faces-of-young-2...
- godelski 4y agoThese are good points! I think another important aspect is that the links you show are only averaging faces of a single gender. The first twitter link to me seems to have features that are both masculine and feminine.