5 ms·

Some of these have designs that are so... specific? holistic? Not sure what I'm shooting for. But, I am wondering if you ran an image similarity search on th

Some of these have designs that are so... specific? holistic? Not sure what I'm shooting for. But, I am wondering if you ran an image similarity search on them vs. the training set would you find matches or are they actually unique? https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed09985.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed09985.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed13901.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed13901.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed31957.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed31957.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed25075.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed25075.jpg

mintplant 6y ago
I know a guy who was developing a reverse image search system in this niche. I'll pass this on to him.
- duskwuff 6y ago
  There are a couple of broad-based "fan art" reverse image search engines out there. One pretty comprehensive one is https://saucenao.com/ https://saucenao.com/.
GauntletWizard 6y ago
I'm convinced that for 90% of these "Does not exist" generators, you can identify two or three hugely influential images, and almost all the rest is noise. There's an image where the hair comes from, an image for the facial structure, and then some perturbations are made to eye color or jawline. Not to say that isn't impressive! Generating convincing fakes, even 1% of the time, even if they're not super unique (There's plenty of people who "look exactly like" in the real world, too!) is a big deal.
- gwern 6y ago
  That is obviously not true if you watch any interpolation videos or do any nearest-neighbor lookups.
dylz 6y ago
It was trained on a tagged art gallery that also contains fan art. I would expect some over prevalence of certain popular characters. For 25075, I would assume https://sonic.fandom.com/wiki/Amy_Rose#Heroes https://sonic.fandom.com/wiki/Amy_Rose#Heroes
- Izkata 6y ago
  For that same one, I see hints of https://sonic.fandom.com/wiki/Rouge_the_Bat https://sonic.fandom.com/wiki/Rouge_the_Bat (to the point where I thought of Rouge before seeing your comment)
gwern 6y ago
You would not. Many GAN papers do nearest-neighbor lookups, and typically, generated samples are clearly different. Since Arfa's model is high-quality and the interps look fine, I would not expect it to be any different. (It would be hard to check because there are no pretrained classifiers to provide an embedding to do the search in, but one could do reverse-encoding.) The fact that nearest-neighbor lookups do not find exact overlaps between the training dataset and a large number of random samples has always been one of the arguments I use against the widespread misconception that GANs 'just memorize' data: https://www.gwern.net/Faces#faq https://www.gwern.net/Faces#faq That so many people are convinced that a given datapoint must be an exact copy of a training datapoint - "it looks exactly like a Zootopia character I recognize!" - is really quite a compliment to the GAN...

mintplant 6y ago

I know a guy who was developing a reverse image search system in this niche. I'll pass this on to him.

duskwuff 6y ago
There are a couple of broad-based "fan art" reverse image search engines out there. One pretty comprehensive one is https://saucenao.com/ https://saucenao.com/.

GauntletWizard 6y ago

I'm convinced that for 90% of these "Does not exist" generators, you can identify two or three hugely influential images, and almost all the rest is noise. There's an image where the hair comes from, an image for the facial structure, and then some perturbations are made to eye color or jawline. Not to say that isn't impressive! Generating convincing fakes, even 1% of the time, even if they're not super unique (There's plenty of people who "look exactly like" in the real world, too!) is a big deal.

gwern 6y ago
That is obviously not true if you watch any interpolation videos or do any nearest-neighbor lookups.

dylz 6y ago

It was trained on a tagged art gallery that also contains fan art. I would expect some over prevalence of certain popular characters. For 25075, I would assume https://sonic.fandom.com/wiki/Amy_Rose#Heroes https://sonic.fandom.com/wiki/Amy_Rose#Heroes

Izkata 6y ago
For that same one, I see hints of https://sonic.fandom.com/wiki/Rouge_the_Bat https://sonic.fandom.com/wiki/Rouge_the_Bat (to the point where I thought of Rouge before seeing your comment)

gwern 6y ago

You would not. Many GAN papers do nearest-neighbor lookups, and typically, generated samples are clearly different. Since Arfa's model is high-quality and the interps look fine, I would not expect it to be any different. (It would be hard to check because there are no pretrained classifiers to provide an embedding to do the search in, but one could do reverse-encoding.) The fact that nearest-neighbor lookups do not find exact overlaps between the training dataset and a large number of random samples has always been one of the arguments I use against the widespread misconception that GANs 'just memorize' data: https://www.gwern.net/Faces#faq https://www.gwern.net/Faces#faq That so many people are convinced that a given datapoint must be an exact copy of a training datapoint - "it looks exactly like a Zootopia character I recognize!" - is really quite a compliment to the GAN...