4 ms·

Air Head by shy kids https://www.youtube.com/watch?v=G4wJ4WeJrz4 https://www.youtube.com/watch?v=G4wJ4WeJrz4 You also don't see the everyday people using it t

Air Head by shy kids https://www.youtube.com/watch?v=G4wJ4WeJrz4 https://www.youtube.com/watch?v=G4wJ4WeJrz4 You also don't see the everyday people using it to generate pictures about their own experiences and lives, using it to populate their DnD worlds, their personal unpublishable fanfics, etc.

YurgenJurgensen 2y ago
...and explain why you think it's good. This part is important since I'm still yet to see evidence that the AI artist is capable of thinking more deeply than "(full_body:1.2) , best quality, (8k, RAW photo, best quality, masterpiece:1.2) , (realistic, photo-realistic:1.4) , ultra-detailed, (Kpop idol) , perfect detail, looking at the viewer, makeup, pretty South Korean lady wearing a bathing suit, wet skin, light reflections, angelic cute face, half body shot, short brownish black hair".
- thaumasiotes 2y ago
  You've gotten me confused about what you're going for. I figure the goal is "this is what this would look like, if there were a person with a balloon for a head and we captured him with a video camera". The linked video isn't there yet, but it seems obvious what would or wouldn't make it good, and the list of traits you describe is really all I want in an image generator. I don't care what the girl is thinking about; that doesn't show up in the image anyway. I do want to be able to say "no, make the hair shorter", or "turn her to face more toward the left of the frame" or "have her pointing at the door in the background". If that command takes the form of a comma-delimited list of traits, so what?
thaumasiotes 2y ago
The balloon is a different shape from shot to shot, and - much worse - in a few of the shots it appears to be hovering in front of the rest of the image rather than occupying the space where Air Head's neck should be. Is it better than I could do? Yep. Is it failing a bar that any human would pass? Yep. Is it better than relevant artists? Nope. But it is cheaper.
- echelon 2y ago
  Literally everyone in this space is working on controllability. The cherry picked issue you cited has hundreds of the best minds tackling it as we speak. This is the worst it will ever look. It's only going to improve from here. Compare that to the early days of silent film. The profess with Gen AI is astounding. We're going to have Disney/Pixar and Scorsese outputs by the end of the decade. (I'd be willing to wager even sooner than that.)
  - thaumasiotes 2y ago
    > The cherry picked issue you cited has hundreds of the best minds tackling it as we speak. Did you mean to respond to my other, nearby comment? The issue I cite above isn't cherry-picked in any sense; it's a big, glaring problem with what was cited as an example of "good work". If someone's head is a balloon, the balloon should occupy the same position in 3D space as the head would. > Compare that to the early days of silent film. This is an interesting comparison. I don't think it really works. Early silent films were aware of what could and couldn't be done in the medium; there aren't any that rely on a nonexistent soundtrack. Generative AI stuff doesn't seem to be very concerned with "what kinds of things can we do well?". Instead, they're attempting everything, almost none of it is being done particularly well, and there are theoretical arguments over whether it makes sense to try to improve on individual tasks.

YurgenJurgensen 2y ago

...and explain why you think it's good. This part is important since I'm still yet to see evidence that the AI artist is capable of thinking more deeply than "(full_body:1.2) , best quality, (8k, RAW photo, best quality, masterpiece:1.2) , (realistic, photo-realistic:1.4) , ultra-detailed, (Kpop idol) , perfect detail, looking at the viewer, makeup, pretty South Korean lady wearing a bathing suit, wet skin, light reflections, angelic cute face, half body shot, short brownish black hair".

thaumasiotes 2y ago
You've gotten me confused about what you're going for. I figure the goal is "this is what this would look like, if there were a person with a balloon for a head and we captured him with a video camera". The linked video isn't there yet, but it seems obvious what would or wouldn't make it good, and the list of traits you describe is really all I want in an image generator. I don't care what the girl is thinking about; that doesn't show up in the image anyway. I do want to be able to say "no, make the hair shorter", or "turn her to face more toward the left of the frame" or "have her pointing at the door in the background". If that command takes the form of a comma-delimited list of traits, so what?

thaumasiotes 2y ago

The balloon is a different shape from shot to shot, and - much worse - in a few of the shots it appears to be hovering in front of the rest of the image rather than occupying the space where Air Head's neck should be. Is it better than I could do? Yep. Is it failing a bar that any human would pass? Yep. Is it better than relevant artists? Nope. But it is cheaper.

echelon 2y ago
Literally everyone in this space is working on controllability. The cherry picked issue you cited has hundreds of the best minds tackling it as we speak. This is the worst it will ever look. It's only going to improve from here. Compare that to the early days of silent film. The profess with Gen AI is astounding. We're going to have Disney/Pixar and Scorsese outputs by the end of the decade. (I'd be willing to wager even sooner than that.)
- thaumasiotes 2y ago
  > The cherry picked issue you cited has hundreds of the best minds tackling it as we speak. Did you mean to respond to my other, nearby comment? The issue I cite above isn't cherry-picked in any sense; it's a big, glaring problem with what was cited as an example of "good work". If someone's head is a balloon, the balloon should occupy the same position in 3D space as the head would. > Compare that to the early days of silent film. This is an interesting comparison. I don't think it really works. Early silent films were aware of what could and couldn't be done in the medium; there aren't any that rely on a nonexistent soundtrack. Generative AI stuff doesn't seem to be very concerned with "what kinds of things can we do well?". Instead, they're attempting everything, almost none of it is being done particularly well, and there are theoretical arguments over whether it makes sense to try to improve on individual tasks.