I think it's quite easy to distinguish b/w DALL-E and human.
The biggest problem w/ current DALL-E is that it's really really bad at details, unless it's a stereotypical one. DALL-E delivers pretty good impression, but almost always fails to fill in the last mile. This is because DALL-E currently doesn't know what "perfection for human" is like..
Also, often times, DALL-E doesn't have a definitive style, meaning it composes images with different styles in one place. This is quite noticeable in the second image for "Teddy bears, mixing sparkling chemicals as mad scientists". The bubbles are obviously not in the same style. This hardly happens for humans because, welp, when someone draws a picture, one person draws with a limited set of tools in a single style. DALL-E itself is not subject to such limitations.
The obvious markers to me is mistakes / skill level. The AI is making mistakes no human would and it’s perfecting advanced techniques only experts pull off.
For example the space ones have incredible style, texture and shading, but the basket ball lines are nonsensical. While the human ones are quite basic and artistically primitive but all of the easy things are flawless.
And then the bears have an eye on their forehead and missing half their mouth mixed with extremely good textures everywhere else.
The human drawn ones were much more interesting to look at, as art, usually because of the emotion and hints of narrative conveyed by facial expressions. I also liked the giant cat.
These days famous artists often don't execute the artwork themselves, they leave that to craftspeople. They have the idea, and the fame to imbue it with attention which connects it with the semantic fabric of the current art scene.
This is mainly because DALLE-2 generates at 64x64, then uses two separate upscalers to reach 256px then 1024px. The problem is that the upscalers are much less capable than the main model, and each upscaling stage tends to compound the mistakes of the upstream image.
I imagine DALLE-3 will generate at 256px directly, and use something like latent diffusion for a free upscale to 1024px. This should eliminate most of the high frequency artifacts.
The best, most concise description of AI generated content (whether textual or visual) I’ve read (I believe it was here on HN) is that “it gets to the topic, but not to the point”.
In other words, while GPT, DALL-E, etc. are good at making things that superficially match the prompt given to them, vaguely seeming coherent, the moment you inspect them closer they do not guide your attention, or create a broader artefact with inter coherent sub parts, the way a human artist/writer would.
It’s hard to tell if that’s an inherent limitation of approaches based on statistical sampling rather than “true understanding” or not. That criticism has held so far over the last decade of progress, so I doubt “more parameters/training data/resolution/etc” is the answer.
A lot of designers just prostitute themselves to making the art as requested - very few infuse it with their own ideas. But they would like to. It won't surprise me if corporations replace them all with AIs. But people will see through it. It won't score as well. A decade or two might pass where almost all communication art is done by AI. In the end, designers will be in high demand because - Marvel movies to the contrary - you can't replace ideas and substance by just visually manipulating stuff. Even if that form of legerdemain works in the short run, it wears off because humans are drawn toward novel, human ideas. This is the whole reason 70s disco collapsed and punk music sprang up. The idea that you can endlessly sell generation after generation complete rip-off garbage runs into a wall because you have to dumb down the population to such an extent that e.g. they're all hooked on fentanyl and can't actually buy anything if you do that. So you then have to retread and market to the people with brains left, but they won't accept recycled corporate trash, and they're the ones you need to keep your brand alive.
You can buy and shovel out all the Maroon 5 you want but you can't be cool unless you co-opt Rage or GnR or some band that's legitimately cool and meaningful to people and outside the reach of what everyone knows was already for sale. AI art is always gonna be corporate garbage regardless of how perfect it looks because the people who guide it will always be corporate douchebags.
It's not that DALL-E2 has a lack of stylistic continuity, or "wholeness" in the image. That's just how it comes across to an untrained eye. That's what engineers are working hard to fix, but it can't fill the uncanny valley.
What you're really perceiving is that every detail in the human work has humor. The humor runs through it; everything, mistakes and all, is shot through with what we visually can see as a sense of the personality of the person who drew it. Not that an AI can't imitate each one, but it can only think inside the conceptual box of what it's trained to imitate.
For instance, the first graphic concept that came to my mind about astronauts playing basketball in space with cats would be to have the whole scene of a cat dunking in space shown in the reflection of an astronaut's visor. The raw literal image is cheap crap without the viewer seeing some human perspective that's trying to be conveyed... in other words, engineers need to go to art school because they're misunderstanding the value proposition behind human-made art.
I think the AI drawings are far better than the human ones.
The human ones are basically a collage of items accurately created to a strict specification, while objects in the AI-drawn ones are all integrated and more artistic and creative.
Although this is probably because the humans were people motivated by money who were paid to spend 15-30 minutes to make the artwork and just tried to check off the requirements (I can't possibly imagine a decent artwork being created in 15-30 minutes).
It's very interesting though that humans are more robotic than neural networks!
Superficially, the AI-drawn images seem better. But I showed these images to my wife asking if she could tell the difference and only the human ones invoked any sort of emotional reaction from her. The human-drawn ones were more interesting visually and told more of a story. She laughed at a few of them. She couldn't really tell the human from AI, but she definitely enjoyed the human-drawn ones more.
The cats not having space suits is a fun detail; it seems to 'understand' clothing on things and some other features now; but it doesn't understand why the astronaut has the space suit or that the cat would obviously need one. I wonder if 5 year olds notice that.
Also Saturn being a basketball is a dead giveaway. dall-e is terrible at localizing stuff to just one part of an image, so it doesn’t know that “space” doesn’t have to be basketball themed as well. It reminds me of when someone tried the prompt “happy sisyphus” and it drew a happy guy carrying a rock with a smiley face on it.
The biggest problem w/ current DALL-E is that it's really really bad at details, unless it's a stereotypical one. DALL-E delivers pretty good impression, but almost always fails to fill in the last mile. This is because DALL-E currently doesn't know what "perfection for human" is like..
Also, often times, DALL-E doesn't have a definitive style, meaning it composes images with different styles in one place. This is quite noticeable in the second image for "Teddy bears, mixing sparkling chemicals as mad scientists". The bubbles are obviously not in the same style. This hardly happens for humans because, welp, when someone draws a picture, one person draws with a limited set of tools in a single style. DALL-E itself is not subject to such limitations.