The generated image of two dice (https://x.com/thomasahle/status/1800258720074490245) was dalle 1 level.

Just randomly sprinkled eyes on the sides. I wonder why they chose to showcase that.

What eyer are you talking about? That’s two “hand-sketched” dice, isn’t it?

Did you look at the eyes/pips?

On the side with 5, they are overlapping. On the side with 4, some of them are half missing. On the side with 3, they are arranged in triangle instead of a straight line.

Not to talk about that 2 and 5 should be on opposing sides, same with 3 and 4.

It's basically like early AI being unable to generate hands, or making 6 fingers.

Yeah, the image generation felt really…cheap?…tasteless? but everything else was really impressive.

Personalization really feels like the missing link here. The images it creates are highly contextual, which increases their value dramatically. Nobody on Reddit wants to see the AI generated T. rex with a tutu on a surfboard, but in a group chat where your dancer buddy Rex is learning to surf, it’s a killer. The image AI can even use photos to learn who a person is. That opens up a ton of cool ways to communicate with friends

> in a group chat where your dancer buddy Rex is learning to surf, it’s a killer.

Maybe, but this class of jokes/riffs is going to get old, fast.

It's what i expected they weren't going to open the pandoras box of realistic photogen on imessage lol, thats why the limit to illustration, cartoon etc, is there to limit the liability of it going wild, they can add more "types" later as they get things more tested, realistically its just prompts hidden behind bubbles, but allows them to slowly roll out options that they've heavily vetted.

I think that basically stretched the limit of what local model can achieve today, which also makes their image API almost useless for any serious generative art developers.

Fwiw I don't think "serious generative art developers" are the target audience at this point, that's probably on the order of .01% of their users