4 ms·

Biggest/best tell for me is not the image but the subject matter. The images with some interesting twist always come from the humans. For instance one image s

Biggest/best tell for me is not the image but the subject matter. The images with some interesting twist always come from the humans. For instance one image shown to me had human hands at a laptop, but they were coming through the screen to type. That is playing with human perception, twisting reality to make an artistic point (the self-typing laptop where the user comes from inside the machine). I've never seen an AI-generated image that projects any artistic depth. They are beautiful and sometimes funny, but never interesting. Of course most human-created images are equally shallow, but if you do see depth then you know it came from a human. Put this test to an art historian/appraiser or architect and they would probably score very well.

hn_throwaway_99 4y ago
This was my "tell" as well. I don't know a ton about DALL-E, but it seems like most of the pictures are in the style of "show X doing Y in the style of Z": "Show a bear carrying a pizza in a realistic style". The composition of the DALL-E ones is almost "random association", e.g. acid-trip type associations. For each image I'd ask myself "What purpose would a human have for creating this?" and it usually helped get me the right answer.
- sandworm101 4y ago
  But what if the human is tasked with creating an image that looks like it was created by a machine? If you gave human artists the same taskings (X and Y in style of Z) then they would also create a bunch of purposeless images. That is the dataset I would create if I wanted to promote my AI project.
  - cowtools 4y ago
    I think that a human's portrayal would sometimes still be noticeably different. For example, if you prompted a human and DALL-E to draw "a fork in an electrical outlet", the human would probably portray the fork as shocking and exploding, whereas DALL-E would probably just portray a fork being put into an electrical outlet and nothing happening. Maybe that's a bad example because there may be enough depictions of forks in electrical outlets for DALL-E to learn to depict them as exploding, but my point is that DALL-E can't reason generally about the composition of the image.
taylodl 4y ago
The human-created images had something interesting to say, whereas the computer-generated images did not. I do find this to be an interesting variation of the Turing Test.

hn_throwaway_99 4y ago

This was my "tell" as well. I don't know a ton about DALL-E, but it seems like most of the pictures are in the style of "show X doing Y in the style of Z": "Show a bear carrying a pizza in a realistic style". The composition of the DALL-E ones is almost "random association", e.g. acid-trip type associations. For each image I'd ask myself "What purpose would a human have for creating this?" and it usually helped get me the right answer.

sandworm101 4y ago
But what if the human is tasked with creating an image that looks like it was created by a machine? If you gave human artists the same taskings (X and Y in style of Z) then they would also create a bunch of purposeless images. That is the dataset I would create if I wanted to promote my AI project.
- cowtools 4y ago
  I think that a human's portrayal would sometimes still be noticeably different. For example, if you prompted a human and DALL-E to draw "a fork in an electrical outlet", the human would probably portray the fork as shocking and exploding, whereas DALL-E would probably just portray a fork being put into an electrical outlet and nothing happening. Maybe that's a bad example because there may be enough depictions of forks in electrical outlets for DALL-E to learn to depict them as exploding, but my point is that DALL-E can't reason generally about the composition of the image.

taylodl 4y ago

The human-created images had something interesting to say, whereas the computer-generated images did not. I do find this to be an interesting variation of the Turing Test.