It’s usually that the AI messes up easy stuff while perfecting hard stuff. A human might struggle to draw a hand, but they don’t massively smudge and distort what should be straight lines. I saw dall e spit out basket balls where the black lines look like spaghetti.
but on the whole most of the output is very good - the images on this website aren't particularly taxing for it and I don't think cherry picking would be needed to replicate this level of quality.
I still largely stand by my point. But there are definitely subjects that Dall-E is worse at and would give the game away. I guess it depends what you mean by cherry-picking - it's certainly not the case that most Dall-E output would be easy to spot - more that it has some specific failure cases.
Not likely because there was quite a lot of work involved by humans. They demonstrated the output of AI and humans for the same prompt. The setup seemed more scientific than the current post (which seems more like a website trying to make a point).
Is it me or is this very depressing and scary? I always assumed the creative industries would be relatively resistant to automation, with more algorithmic jobs more easily replaceable. But with this you never need to buy stock pictures ever again, right? I saw several stock-picture-esque images of indistinguishable quality, and yet with incredible specificity to the user-supplied description.
I thought this way originally but I have realised these tools are being used right now to massively amplify the creative output of individuals. AI text to voice tech is being used for a lot of hilarious content right now. I saw an image upscailer ai used in a 1 week project to rebuild a ps2 game to make it look modern.
There is a long runway of human guided ai generation. Imagine if a single human was able to produce an entire animated TV show and all it took was good story telling skills without requiring 100 animators.
This is the right way to think about this IMO. Wouldn’t it be amazing if tools like this massively increased the scope of what is possible for a single artist?
Something I think about a lot is video game creation - what if AI tools let you quickly create 3D objects and a world without a big, dedicated team of artists?
Humorously, the narration on that video also sounds like an AI. The rising inflection on nearly every clause almost sounds like they’re recording every sentence separately
It’s similar to how photography grants anyone the technical ability to render a photo-realistic image, but the majority of photos we see are easily discarded from memory. It did put painters out of the portrait business, but also freed them from a realist ideal.
I find it amazing and exciting - but then I don't earn my living as an illustrator.
I suspect there was similar wailings and gnashings of teeth when photography was invented. Humans adapt and it's us that decides what to value. We are quite fond of other people so tastes will alter to preserve that balance.
There was a clear career path for anyone out of a job due to photographs though. Become a photographer. Which was a perfectly fine career for a very long time. It still is, really, despite the high quality cameras available to many people, because it requires artistic and technical skill.
The career path for an artist trying to find something else while AI is developing at this pace is...less clear, and definitely scarier.
Photography didn't replace visual art though (which is just as well for painters and sculptors really, the skill needed to operate an early camera was pretty orthogonal). A subset doing the grunt work of simply recording details for prosperity had a competitive alternative, but even in the subset of image production where photorealism is considered desirable, people still commission painters for "lifelike" portraits and Canaletto renderings of Venice are still worth a fortune despite a near infinitude of similar images on Instagram.
Same goes for this really, but perhaps even more so. It's impressive at producing eye-catching photos quality images of simple concepts you might use in advertising campaigns, and can do a good job of reproducing well defined cartoon styles. But the artists working in those sort of roles are already working digitally and using image libraries, and this is a massive productivity boost for them - dozens of "$brand on beach" concept images for no effort at all, and they'll still need all the retouching and layering and colour balancing and combining with other images. And as soon as the brief gets more precise about details of layout and colour scheme or what they want the model to look like, they're either back to their original creative flow or designing text prompts and curating images for the brief is an advanced skill in itself. Non artists can produce "good enough" images too (looks great for memes), but much like non-artists wielding cameras they're mostly satisfying their own whims rather than competing with the business of art
Creative industries are still safe from automation. This doesn't do anything except generate nice stock art or album covers. Unless you're in those industries, there's nothing to worry about. It can't even do text.
The quality is not indistinguishable to any deliberate observation. Plus, I think people will begin to casually notice how off it is once they become ‘trained’ as to the quirks of these images. That is, unless the quality of the generated images continues to steadily improve.
Every time I think I have come to understand the limits of ML generated content, I see something a few months later that completely surpasses what I believed possible.
From the videos I have seen of it in use, there is no way the prompts had existing material. One prompt was “painting of a goat in the style of the Mona Lisa taking photos with an iPad”. And it spat out 10 images which showed exactly that.
DALL-E 2 does support 'inpainting' and 'uncrop' (zoom out, synthesizing edges) functionality which can be used to edit AI or human-generated images. I don't think OP used any (at least wittingly): one of the ones I looked at in OP were ones I recalled as inpainted/uncropped, those are relatively unusual compared to pure generation, and it'd be a bit unfair to include any.
I'm slightly surprised to report a 100% score. However, I think I'm actually recognising DALL-E-2's _style_, and not "this picture is CG". It feels more like I'm identifying the work of an artist I'm very familiar with, rather than identifying fundamentally different categories of creator.
I think all it would take to "defeat" me here would be greater variety of high-quality image generation AIs.
I expect in a few years that will be the case, unless the "style" of DALL-E-2 is something fundamental, instead of an artifact of its specific implementation.
Even now just selecting Dalle propmpts that output things that look like basic ray tracer tests will probably make it undecidable for anyone not bothering to recrating the scene and chech that it's physically coherent.
well, Theres still plenty telltale signs of AI handwork here. Quasi-faces on places no human would place them or would articulate a detail much better. edge coherence is also very important. Seeing smudges of paint where it doesn't belong at all, lines no human would be able to make and errors only humans make but where AI struggles with.
My point is that Dalle can do things like the Cornell box pretty much perfectly, so if you want to make a test where it's impossible to decide who made it you should sample those type of prompts.
If you open up the image in a new tab to see the full resolution and look at fine details it's very easy to spot AI ones - AI seems to especially struggle on reflective details, it obviously doesn't know it's meant to be a reflection, so it gets details subtly wrong or weird in the mirror image.
Also anything mechanical is just way off on AI generated things - a car with the fuel door in the middle of the passenger door, a missing tire but you see no axle and nothing makes mechanical sense.
If you flip a coin 30 times you won't get 15 heads and 15 tails. But getting 30 heads in a row would be one in a billion (literally); HN is popular but I doubt there's a billion people going to the site...
Right, but there is exactly one outcome with all heads, and there are at least millions of combinations with 50% heads and 50% tails.
To make it more obvious, let's take just 4 flips. There is exactly one combination with 4 heads:
HHHH
Whereas all the following will give you 2 heads and 2 tails:
HHTT
HTHT
HTTH
THHT
THTH
TTHH
So you're 6x more likely to get 50% heads/tails than 100% heads. Scale that up to 30; the chances of getting 15 heads and 15 tails is not the most common thing but not unusual. The chances of getting all heads is quite rare -- if every human on earth did this test, we'd expect around 7 such instances total.
Having already seen a few DALL-E generated images, there seemed to features I could pick up on that would allow me to identify them such as proportions, composition or unnecessary random detail. For example, in one image, there was a trail from what looked like a spaceship in the sky but the spaceship itself was obscured by a face in the foreground. It was hard to believe any human artist would choose that composition as it's quite jarring to look at.
I got 26/30, but only because I knew of DALL-E 2's "style", i.e. the robot ones were mostly drawings. if they gave a more "photorealistic" images, I could have been thrown off easily I think
Also - please speed up the delay switching to a new image! I bet this would feel way better if the delay was in the 0-0.5s range instead of multiple seconds. (I bet it’s some runaway JS, investigating…)
Edit: `changeImage` is blocked by `await saveDB`, and the POST is 500ing. (e.g. uBlock makes that AWS request fail fast – why desktop people are less affected)
Maybe use `Promise.allSettled` to make them parallel, and prepare the data being sent beforehand to avoid state race conditions :) thanks for leaving sourcemaps on!
> Also - please speed up the delay switching to a new image!
I came here to suggest the opposite, as I didn't have time to read the full prompt which generated the image, when it was a generated picture.
So I guess the conclusion is instead: allow users to click a second time on any of the buttons/picture to proceed at their own leisure, so the ones who want to read can read, and the ones who want to speed through it, can do that too.
Nice, that’s even better :) add some keyboard controls (maybe left for artist, right for robot?, and then either key to advance) and folks can zoom through!
I wonder how hard the jump to 3d models would be. It would be great if we had an AI thatyou can just prompt to generate a high quality 3d model. I know there are already Models that can create 3d models from a collection of images from different angles so I think it could probably be possible.
26/30. Since 90%+ of images seem to be have been generated by a machine, a better challenge would be to pick which image was generated by a human from a gallery of images.
If you knew ahead of time that it was 50/50, then you could adjust your odds as you went on; i.e., on image 28, if there had been 14 CG ones and 13 human ones, you'd know a priori it had a 2/3 chance of being human.
It would be nice if they told you in more detail how the algorithm worked. How well curated are the AI and human images? It seems like selecting the absolute best AI ones and the worst human ones would make it easy to make things look better than they really are.
A genuinely "fair" way to do it I think would be this:
1. Randomly collect a few thousand images from DeviantArt
2. Ask people to write a short description of each image, similar to the prompts given to DALL-E
3. Generate DALL-E images from all the prompts
Randomly show a human or AI image with 50% probability.
That should give you a good measure of the actual state of the art.
One thing I noticed is that the robot is not very good at composition. Awkward crops can be a bit of a giveaway. I suspect this could an artefact of the training process, where the images were cropped and shifted randomly (speculation).
Perhaps giving the "real" art a random crop would level the playing field slightly.
27/30. You can tell the AI ones because they don't look like they were made with the same constraints as a human; they merge, distort, twist together lines, shapes, and colors, rather than have images split into the compartmentalized pieces that humans like to do. I find the best place to check for AI art is to look at the overlaps; this is how I figured out the kindergartener one, for example: there were no obvious "AI-like" patterns except for the one section where the drawn lady's legs intersected with the ground line, and you could see that DALL-E 2 didn't know what to do, putting one leg over the line and one underneath, which can't be done unless one leg was drawn, then the ground, then the other, which is not what a human does.
I scored 29/30, only missing one because there were so many machine generated ones I started to just click the machine button repeatedly. Then I checked more closely and got all the rest right.
The machine generated ones are usually easy to spot if you look closely, because parts of the image are just nonsense (e.g. weird patterns, things out of place, odd focus). Machine generated images clearly have no actual understanding of the real world.
135 comments
[ 3.1 ms ] story [ 232 ms ] threadhttps://unsplash.com/license
There's always a few duds and some definite limits (https://www.lesswrong.com/posts/uKp6tBFStnsvrot5t/what-dall-... )
but on the whole most of the output is very good - the images on this website aren't particularly taxing for it and I don't think cherry picking would be needed to replicate this level of quality.
I think the current post really gives the wrong impression.
Maybe it was cherry picked?
There is a long runway of human guided ai generation. Imagine if a single human was able to produce an entire animated TV show and all it took was good story telling skills without requiring 100 animators.
Something I think about a lot is video game creation - what if AI tools let you quickly create 3D objects and a world without a big, dedicated team of artists?
(Check this out by Nvidia! https://m.youtube.com/watch?v=5j8I7V6blqM)
I suspect there was similar wailings and gnashings of teeth when photography was invented. Humans adapt and it's us that decides what to value. We are quite fond of other people so tastes will alter to preserve that balance.
The career path for an artist trying to find something else while AI is developing at this pace is...less clear, and definitely scarier.
Same goes for this really, but perhaps even more so. It's impressive at producing eye-catching photos quality images of simple concepts you might use in advertising campaigns, and can do a good job of reproducing well defined cartoon styles. But the artists working in those sort of roles are already working digitally and using image libraries, and this is a massive productivity boost for them - dozens of "$brand on beach" concept images for no effort at all, and they'll still need all the retouching and layering and colour balancing and combining with other images. And as soon as the brief gets more precise about details of layout and colour scheme or what they want the model to look like, they're either back to their original creative flow or designing text prompts and curating images for the brief is an advanced skill in itself. Non artists can produce "good enough" images too (looks great for memes), but much like non-artists wielding cameras they're mostly satisfying their own whims rather than competing with the business of art
An unpaid intern adding text to AI generated images would be trivial.
I think all it would take to "defeat" me here would be greater variety of high-quality image generation AIs.
I expect in a few years that will be the case, unless the "style" of DALL-E-2 is something fundamental, instead of an artifact of its specific implementation.
Also anything mechanical is just way off on AI generated things - a car with the fuel door in the middle of the passenger door, a missing tire but you see no axle and nothing makes mechanical sense.
https://lucid.fandom.com/wiki/Reality_check
I can see DALL-E-2 stuff spreading around WhatsApp groups like wildfire, masquerading as something else
HHHHHHHHHHHHHHHHHHHHHHHHHHHHHH
and
HHTHTHTTTTHTHTHTHTHTHHHTHTHTHTHTH
are equally as likely outcomes.
To make it more obvious, let's take just 4 flips. There is exactly one combination with 4 heads:
HHHH
Whereas all the following will give you 2 heads and 2 tails:
HHTT HTHT HTTH THHT THTH TTHH
So you're 6x more likely to get 50% heads/tails than 100% heads. Scale that up to 30; the chances of getting 15 heads and 15 tails is not the most common thing but not unusual. The chances of getting all heads is quite rare -- if every human on earth did this test, we'd expect around 7 such instances total.
Also - please speed up the delay switching to a new image! I bet this would feel way better if the delay was in the 0-0.5s range instead of multiple seconds. (I bet it’s some runaway JS, investigating…)
Edit: `changeImage` is blocked by `await saveDB`, and the POST is 500ing. (e.g. uBlock makes that AWS request fail fast – why desktop people are less affected)
Maybe use `Promise.allSettled` to make them parallel, and prepare the data being sent beforehand to avoid state race conditions :) thanks for leaving sourcemaps on!
I came here to suggest the opposite, as I didn't have time to read the full prompt which generated the image, when it was a generated picture.
So I guess the conclusion is instead: allow users to click a second time on any of the buttons/picture to proceed at their own leisure, so the ones who want to read can read, and the ones who want to speed through it, can do that too.
I struggled with some of the stylised paintings, but it's clear that DALL-E cannot do detail.
[1] https://thisimagedoesnotexist.com/images/130.jpeg
It would be nice if they told you in more detail how the algorithm worked. How well curated are the AI and human images? It seems like selecting the absolute best AI ones and the worst human ones would make it easy to make things look better than they really are.
A genuinely "fair" way to do it I think would be this:
1. Randomly collect a few thousand images from DeviantArt
2. Ask people to write a short description of each image, similar to the prompts given to DALL-E
3. Generate DALL-E images from all the prompts
Randomly show a human or AI image with 50% probability.
That should give you a good measure of the actual state of the art.
Perhaps giving the "real" art a random crop would level the playing field slightly.
The machine generated ones are usually easy to spot if you look closely, because parts of the image are just nonsense (e.g. weird patterns, things out of place, odd focus). Machine generated images clearly have no actual understanding of the real world.
https://imgur.com/a/hF6R5Sx