4 ms·

Not really, because you cannot efficiently search by face and body features. IMHO publishing an image that looks like a famous actress, without indicating which

by js8 3y ago

Not really, because you cannot efficiently search by face and body features. IMHO publishing an image that looks like a famous actress, without indicating which actress it is, should be allowed. Sure, perhaps somebody could do an image search, but that's already more complicated process than synthesize your own fantasy. There are other edge cases, for example, should an image of "Princess Leia" be allowed to be shared under that title, when it implies the resemblance to the actress? I think the answer is probably yes, like with the body features, character name is not enough to identify a person. But it would not be allowed to mention Carrie Fisher in conjuction with such a picture, even by a 3rd party reposting the image. Edit: Maybe anybody can recognize Emma Watson (myself, I am not sure, not HP fan), but the issue is how you then use any human image generator without potentially infringing some random actress (effectively banning them)? There are definitely facial features I like and I cannot discern that I like them because I saw them in some actress whose name I have long since forgotten. The actresses tend to have wide variety of appearances, and there is not that many faces. It's gonna be a very contested rule. That's why I think it would be much easier to simply go after the name connection with the actress, rather than facial or bodily features.

pjc50 3y ago
> you cannot efficiently search by face and body features Yet. That feels like something an LLM trained on lots of images would be good at. It's also the basis of biometric identification. (Heck, google lens will give you celebrity identification about 10% of the time even though it's not designed for that. The Bing image search will try to identify the clothing and handbag in the photo instead so it can sell them to you)
- js8 3y ago
  I think I addressed that in the next sentence. Doing an image search using a real image and deciding whether the fake image you find is good enough for your fantasy is already more complicated process than getting two images (one of your fantasy and the real one with the actress) and having AI synthesize them together. Or maybe you need to be more specific how do you imagine people would use the technology.
  - Hizonner 3y ago
    Get a normal, fully-clothed, completely SFW, maybe face-only shot of your target. Paste it into facial-image-similarity search engine that also supports constraining the output using text, and make the text "hardcore blah blah blah". If, as you assume, there are in fact hardcore blah blah blah images out there that look like they're of the target, the image search will find them. A search engine that does something like that doesn't seem too hard to build and will probably come along sometime... especially if the availability of images that you can't search for by pure text has created demand. In fact, it doesn't seem too farfetched for somebody to build a purely text-prompted image searcher that can accept "looks like $target" as part of the search string, and have it go find a picture of the face, then combine that with the rest of the query text. Again, if the images are out there, it'll find them. Not only do the images being out there create demand for the search system, but the search system then creates demand for people to compete in producing and publishing the "best" images. So you might very well find that the skill level for finding the images was even lower than the skill level for prompting them. And of course the shared images would automatically be visible to the target, as would a lot of link sharing and commmenting activity. Pretty much any publicly visible woman, and many publicly visible men, would end up being able to search for their own name and find a bunch of people sharing and talking about obvious fake images of them. Sharing the images really does have different consequences than just generating them, even if you don't put names on them. Also, I don't think there's any functional difference between allowing an image to be tagged as "Princess Leia" or "Hermione Granger" and allowing that image to be tagged as "Carrie Fisher" or "Emma Watson". That just means that the character name becomes a widely known, thinly veiled code for the actress name. I don't know what to do here, but I'm pretty convinced the particular line you want to draw doesn't make sense.
    - js8 3y ago
      > In fact, it doesn't seem too farfetched for somebody to build a purely text-prompted image searcher that can accept "looks like $target" as part of the search string, and have it go find a picture of the face, then combine that with the rest of the query text. Again, if the images are out there, it'll find them. Then it's the search engine which makes/keeps and publishes the association between actress names and their images, and publishing that association would be illegal. In other words, the search engine should not make an inference, this image looks like a specific actress => this image is of this actress. > Also, I don't think there's any functional difference between allowing an image to be tagged as "Princess Leia" or "Hermione Granger" and allowing that image to be tagged as "Carrie Fisher" or "Emma Watson". That just means that the character name becomes a widely known, thinly veiled code for the actress name. I think there is a difference. People already understand that Lena Headey is not an actual IRL evil queen. Look at where the reputation risk comes from - it comes from the association with our real-world identities, not with those of their characters. Why should the actress care if her character is depicted doing something bad? Ultimately I think, the whole "reputation risk" is stupid, and humans should know better not to judge others based on some random image, and consequently, not to worry about being judged from some random image; that would really solve the problem. But perhaps it's too much to ask, so we have to draw the line somewhere.

by js8 3y ago

pjc50 3y ago

> you cannot efficiently search by face and body features Yet. That feels like something an LLM trained on lots of images would be good at. It's also the basis of biometric identification. (Heck, google lens will give you celebrity identification about 10% of the time even though it's not designed for that. The Bing image search will try to identify the clothing and handbag in the photo instead so it can sell them to you)

js8 3y ago
I think I addressed that in the next sentence. Doing an image search using a real image and deciding whether the fake image you find is good enough for your fantasy is already more complicated process than getting two images (one of your fantasy and the real one with the actress) and having AI synthesize them together. Or maybe you need to be more specific how do you imagine people would use the technology.
- Hizonner 3y ago
  Get a normal, fully-clothed, completely SFW, maybe face-only shot of your target. Paste it into facial-image-similarity search engine that also supports constraining the output using text, and make the text "hardcore blah blah blah". If, as you assume, there are in fact hardcore blah blah blah images out there that look like they're of the target, the image search will find them. A search engine that does something like that doesn't seem too hard to build and will probably come along sometime... especially if the availability of images that you can't search for by pure text has created demand. In fact, it doesn't seem too farfetched for somebody to build a purely text-prompted image searcher that can accept "looks like $target" as part of the search string, and have it go find a picture of the face, then combine that with the rest of the query text. Again, if the images are out there, it'll find them. Not only do the images being out there create demand for the search system, but the search system then creates demand for people to compete in producing and publishing the "best" images. So you might very well find that the skill level for finding the images was even lower than the skill level for prompting them. And of course the shared images would automatically be visible to the target, as would a lot of link sharing and commmenting activity. Pretty much any publicly visible woman, and many publicly visible men, would end up being able to search for their own name and find a bunch of people sharing and talking about obvious fake images of them. Sharing the images really does have different consequences than just generating them, even if you don't put names on them. Also, I don't think there's any functional difference between allowing an image to be tagged as "Princess Leia" or "Hermione Granger" and allowing that image to be tagged as "Carrie Fisher" or "Emma Watson". That just means that the character name becomes a widely known, thinly veiled code for the actress name. I don't know what to do here, but I'm pretty convinced the particular line you want to draw doesn't make sense.
  - js8 3y ago
    > In fact, it doesn't seem too farfetched for somebody to build a purely text-prompted image searcher that can accept "looks like $target" as part of the search string, and have it go find a picture of the face, then combine that with the rest of the query text. Again, if the images are out there, it'll find them. Then it's the search engine which makes/keeps and publishes the association between actress names and their images, and publishing that association would be illegal. In other words, the search engine should not make an inference, this image looks like a specific actress => this image is of this actress. > Also, I don't think there's any functional difference between allowing an image to be tagged as "Princess Leia" or "Hermione Granger" and allowing that image to be tagged as "Carrie Fisher" or "Emma Watson". That just means that the character name becomes a widely known, thinly veiled code for the actress name. I think there is a difference. People already understand that Lena Headey is not an actual IRL evil queen. Look at where the reputation risk comes from - it comes from the association with our real-world identities, not with those of their characters. Why should the actress care if her character is depicted doing something bad? Ultimately I think, the whole "reputation risk" is stupid, and humans should know better not to judge others based on some random image, and consequently, not to worry about being judged from some random image; that would really solve the problem. But perhaps it's too much to ask, so we have to draw the line somewhere.