It's fun to play around with it, but like the author found, what you get is often strange or useless. I also find 1k images too small to do much with but I realize making 4k images would be cost prohibitive. I also wish it could generate vector images as well as pixel images. That would be fun to use.
“In working with DALL·E 2, it’s important to be specific about what you want without over-stuffing or adding redundant words.”
I found this to be the most important point from this piece. Often people don't really know what they really want when it comes to creative work, let alone to some omniscient algorithm. In spite of that, it's a delight to see something you love from an unspecific prompt that you won't find with anything you receive from a human.
Dall.E 2 never ceases to amaze me.
For anyone interested in learning about what Dall.E 2 can do, the author also links to the Dall.E 2 prompt book (discussed in this post https://news.ycombinator.com/item?id=32322329).
I want to see a few iterations of describing an image with AI, generating it, describing it again, generating it... Like when passing a piece of text through Google translate back and forth.
There was a tool that could find the "equilibrium" called Translation Party. I don't think it works anymore. I'd love to see one that goes back and forth between DALL-E and an image description algorithm.
According to internet popular belief, you'd end up with a picture of a certain ignominious dictator that unfortunately destroyed Europe in the 1940's. [1]
The fact that it's a derivative of an existing work is noteworthy, but I gave it absolutely no guidance on the topic. If i suggest something it will give it a go with similar fervor. eg https://imgur.com/a/N1qWaSV
Here is "llama in a jersey dunking a basketball like Michael Jordan, shot from below, tilted frame, 35°, Dutch angle, extreme long shot, high detail, dramatic backlighting, epic, digital art": https://imgur.com/a/7LoAtRx
Here is "Llama in a jersey dunking a basketball like Michael Jordan, screenshots from the Miyazaki anime movie", much worst: https://imgur.com/a/g99G7Bn
Craiyon did step up a lot in its understanding recently. The image quality is still not the best but it if you ignore the blurriness, the scary faces, and the weird shapes, it can sometimes be better than dall.e.
Wow. Those models, particularly Imagen, are of an entirely separate calibre. None of that psychedelic foggy memory swirl that is characteristic of the space. I can see why Google Research is hesitant to release them.
After using all of the different models extensively, Stable Diffusion is currently state-of-the-art.
Images are more artistic and less clip art-like than Dall-E, but also don’t have a house style like Midjourney. It’s stunningly good - and open source.
What’s really cool is that the devs have worked hard to optimise the model, so after being trained on 1000 A100s it’ll run happily on an 8gb graphics card or M2 Mac.
I picture in a few years we will be playing around with a code generation tool, and people will be drawing similar conclusions. "You have to be really specific about what you like. If you just say 'chat tool', it will allow you to chat to one other person only."
My current move is creating initial versions of images with Midjourney, which seems to be a bit more "free-spirited" (read: less _literal_, more flexible) and then using DALL-E's replace tool to fill in the weird looking bits. It works pretty well, but it's a multi-step process and requires you have pay for Midjourney and DALL-E.
-- hat pic are playing with "variations" mode - the prompt was: “portrait photo, california beach with female model wearing hat and sunglasses, studio, lens flare, colourful, 4k, high definition, 35mm, HD” --
> it was difficult to find images where the entire llama fit within the frame
I had the same trouble. In my experiment I wanted to generate a Porco Rosso style seaplane. illustration. Sadly none of the generated pictured had the whole of the airplane in them. The wingtips or the tail always got left off.
I found this method to be a reliable workaround: I have downloaded the image I liked the most. Used an image editing software to extend the image in the direction I wanted it to be extended and filled the new area with a solid colour. Cropped a 1024x1024 size rectangle such that it had about 40% generated image, and 60% solid colour. Uploaded the new image and asked DALL-E to infill the solid area while leaving the previously generated area unchanged. Selected from the generated extensions the one I liked the best, downloaded it and merged it with the rest of the picture. Repeated the process as required.
You need a generous amount of overlap so the network can figure out which parts is already there and how best to fit the rest. It's a good idea to look at the image segment you need to be infilled. If you as a human can't figure out what it is you are seeing, then the machine won't be able to figure it out either. It will generate something, but it will look out of context once merged.
The other trick I found: I wanted to make my picture a canvas print, and thus I needed a higher resolution image. Higher even then what I can reasonably hope with the above extension trick. What I did is that I have upscaled the image (used bigjpg.com, but there might be better solutions out there.) After that I had a big image, but of course there weren't many small scale details now on it. So I have sliced it up to 1024x1024 rectangles, uploaded the rectangles to DALL-E and asked it to keep the borders intact but redraw the interior of them. This second trick worked particularly well on an area of the picture which shown a city under the airplane. It has added nice small details like windows and doors and roofs with texture without disturbing the overall composition.
Anytime! I have uploaded the image in question: the initial prompt with first generated images, the extended raw image, and then the one with the added details on the city.
Good question! All of them had the same postfix ", studio ghibli, Hayao Miyazaki, in the style of Porco Rosso, steampunk". I used this for all the generations in the hopes of anchoring the style.
With the prefix of the prompt I described the image. I started the extension operations with "red seaplane over fantasy mediterranean city" but then I quickly realised that this was making the network generate floating cities in the sky for me. :D So then I varied the prompt. "red seaplane on blue sky" in the upper regions and "fantasy mediterranean city" in the lower ones.
I went even more specific and used "mediterranean sea port, stone bridge with arches" prefix for a particular detail where I wanted to retain the bridge (which I liked) but improve on the arches. (which looked quite dingy)
(I have just counted and it seems I have used 27 generations for this one project.)
I was testing to see how close I could get to replicating a t-shirt graphic concept I saw.
I had been using ~"A telephoto shot of A neglected police car from the 1980s Viewed from a 3/4 angle sits in the distance. The entire vehicle is visible but it is overgrown with grass and flowery vines"
This process sounds great, though it seems like DALLE needs to offer tools to do this automagically.
These are trained with pairs of image and caption text, so they work better with text inputs that resemble description for paintings than with simple descriptions or with William-Gibsonian hyperspecified description-text, though it's tempting to do the latter two.
I think "fitting the entire X within the image" is not done on purpose. The results are more aesthetically pleasing when the subject is large, even if a part of it is missing.
Yesterday I saw one of Gandalf eating samples at Costco. I was laughing hysterically for a minute. AI is not supposed to have a sense of humor. That was supposed to be the last province of the human, but it is quite awhile since a human made me laugh like that.
If I write a Python script that cuts together a bunch of pictures and the output makes you laught the script hardly deserves all the credit. It's us humans that create meaning.
And this AI doesn't. Your anecdote is totally unrelated to the idea of AGI in the gp post. The fact that it made you laugh is a happenstance. It was not "trying" to make you laugh.
It’s only unrelated if there’s no proto-AGI going on. Many images give me a moment of doubt, even though I absolutely know that I’m looking at nothing more than the output of a pile of model weights, says I the pile of neurons.
I saw that on reddit. The face was horrific and not at all human like. It didn’t have a sense of humour - it just took a prompt and mashed some things together, but the prompt was funny and the image was horrifying. Not even uncanny valley shit, but “Gandalf was in a bad motorcycle and will never look like a human again” bad.
This tells us little about AGI. It might seem like it does but this is an incredibly narrow specific set of technologies. They work together to produce some startling results (with many limitations) but this is just another narrow application.
I suspect AGI, depending on how its defined, will be with us in some form in the next few decades at most. Just a hunch. This is nothing to do with that mission though imho. Maybe you can read into it something like, "we are solving lots of discrete problems like this, maybe we can somehow glue them together into a higher level program"? That might give you something AI-esque? My guess is that 'true' AGI will have an elegant solution rather than a big bag of stuff glued together.
Machine learning just glues together existing things, which is how art is created. As amusing these pictures are, it's us humans who bring meaning to them, both when producing what these algorithms use as input and when consuming their output. We are the actual magic behind DALL-E.
An AGI wouldn't need us to this extent, or at all. An AGI would also be able to come up with new ways to represent ideas, even ways that are foreign to us.
This is really good fun, actually. Spent some time fucking around with it and it can make some impressive photorealistic stuff like "hoverbus in san francisco by the ferry building, digital photo".
I mostly use it and Midjourney for material for my DnD campaign, but I'm going to need to do a little more work to make the whole thing coherent. Only tried it once and it was okay.
The interesting part is that it can do things like "female ice giant" reasonably whereas google will just give you sexy bikini ice giant for stuff like that which is not the vibe of my campaign!
I ran into this too. When I got my invite, I told a friend I would learn how to talk to DALL-E by having it make some concept art for the game he was designing. I ran through all of my free credits, and most of the first $15 bucket and never really got anything usable.
Even when I re-used the exact prompts from the DALL-E Prompt Book, I didn't get anything near the level of quality and fidelity to the prompt that their examples did.
I know it's not a scam, because it's clearly doing amazing stuff under the hood, but I went away thinking that it wasn't as miraculous as it was claimed to be.
I suspect that many of the "impressive" examples that we see from tools like this have been carefully selected by human curators. I'm sure it's not at the level of "monkeys + typewriters = Shakespeare [if you're sufficiently selective]", but the general idea is still applicable.
Most of DALL-E2 output is great out of the box, the selection process is just fine tuning the results to create something the human in front of the computer likes. DALL-E2 can't mindread, so the image produced might not match what the human had in mind.
There is however one thing to be aware of, the titles posted on /r/dalle2/ and other places are often not the prompts that DALL-E2 got. Instead they are a fun description of the image done by a human after the fact. Random example:
Which is quite a bit less impressive, as the actual prompt doesn't really match the image very well. And if you put "Chased by an amongus segway" into DALL-E2, you won't get an image of that quality either.
I wouldn't at all agree that most Dall-E output is great out of the box. It has areas that it's good at and areas it's poor at.
Here's a result for the prompt of "Woman with green skin, leaves instead of hair, wearing a simple dress, far shot, digital art, hyper-realistic, 8k, ultrahd," for example (all four images)
You will note that none of them are even basically fulfilling the prompt, as well as all four being, in my estimation, ugly and uninteresting. That's not unusual for prompts that involve some element of the fantastic -- though there are corners of less-realistic digital art that it does do well.
You will still note that none of them are far shots, that no depicted character actually has fully green skin, and one of the four has nothing even remotely like leaves for hair. I mean, is it better? Sure. They're less ugly, though none of them are what I'd call great results. But they also aren't really doing a basically competent job of fulfilling the prompt, much less producing a particularly striking or interesting images.
And my point is, outside of a few areas, this is what you get from Dall-E. Lots of misses, and if you're willing to put time into it and work on your results, a few hits. Don't get me wrong, I've gotten stuff from Dall-E that I think is great (I really like this "watercolor painting" for example: https://labs.openai.com/s/AQ7Wy5VHBWcLL5bJ5LbU5SuW) but I think it misrepresents Dall-E to suggest that most of the time it produces basically good images.
I'd say more like, "If you put time and attention into learning its quirks, in its best areas, it'll produce like one in ten images that are basically good."
And, I mean, on some level that's incredible. You can produce 10 images in about three minutes in Dall-E and get some great stuff. But I think people mostly see the top 10% of what Dall-E produces.
I think Dalle's ability to produce good images out of the gate is pretty limited, but I've found that using the fill-in feature along with existing images from google and photoshop, I can pretty much get anything I conceptualize with about 20 minutes of work and like 10 prompts.
It's not fully removing humans from the equation, but you can take something that used to take days and make it a 20 minute operation.
If you're not tired of the whole affair, you should try MidJourney. It's good at different things from DALL-E, but I do feel it produces higher quality pictures on average.
A lot of these posts showing up on HN. I wonder - is it because it is so new, or is it because the ways in which we are to use this technology are so nascent that we are discovering how to use it more precisely daily?
I believe it’s for a few reasons. First, it is jaw dropping incredible for most people in tech who have at least a hint of how most ML works. Second, the AI image generation field is racing ahead, in academics and new trained models, so there’s lots of new news. Thirdly some really great models like Dall-e have been opened for wider access and lots of everyday users are discovering its capabilities and doing blog write-up’s which are not news, but are surely interesting to most.
I've tried out a couple of prompts from the post in Stable Diffusion and as expected the results were much weaker. It has drawn some alpacas and basketballs with little relation between the objects.
I've been playing with Stable Diffusion a lot, and in my experience its results are much weaker then what's shown in this post. The artistic pictures that it generates are beautiful, often more beautiful then Dalle-2 ones. But it has a real problem understanding the basic concepts of anything that is not the simplest task like "draw a character in this or that style". And explaining the situations in detail doesn't help - the AI just stumbles upon basic requests.
Seems like Stable Diffusion has a much more shallow understanding of what it draws and can only produce good result for things very similar to the images it learned from.
For example, it could generate really good dutch still life paintings for me - with fruits, bottles and all the regular expected objects for this genre of painting. But when I've asked it to add some unusual objects to the painting (like a Nintendo switch, or a laptop) - it couldn't grasp this concept and just added more warbled fruit. Even though the system definitely knows how a Switch looks like.
The results in the post are much more impressive. I doubt that Dalle-2 saw a lot of similar images in training, but in all of the styles and examples it definitely understood how a llama would interact with a basketball, what are their relative sizes and stuff like that. On surface results from different engines might look similar, but to me this is an enormous difference in quality and sophistication.
Stable Diffusion has a smaller text encoder than Dalle 2 and other models (Imagen, Parti, Craiyon) so that it can fit into consumer GPUs. I believe StabilityAI will train models based on a larger text encoder, the text encoder is frozen and does not require training, so scaling the text encoder is quite free.
For now this is the biggest bottleneck with Stable Diffusion, the generator is really good and the image quality alone is incredible (managing to outperform Dalle 2 most of the time).
I'm not sure why anyone bothers. StyleGAN2 profile photos are literally all over social media and they're good enough to fool the human reviewers every time I report them.
Hi, author here - that's a great point. When I first saw those results and how inaccurate they were, I thought there was a chance it was returning me an overfitted actual input image from training. Most likely not, but they were so realistic (and I was used to just seeing llamas until this point), that I thought I'd play it safe.
I suspect this is a joke, but I did find that it was a little overzealous with the filtering. I was trying to get someone (not a specific person) shouting or with an angry expression, and a few prompts I came up with were blocked. Not banned though.
I kept getting a scene with "two people holding hands" blocked, it allowed "two people kissing" and then when I tried "and wife" instead of "two people" it banned me. (They unbanned me when I emailed them though.)
Oddly, the ones it blocked were more sfw than several others it allowed, but of course I don’t know what the outputs would’ve been…
At least 10% of web dev today is being good at search prompts for Google. (And that's not necessarily a bad thing, it's just about finding the right tool or pattern for your specific problem)
Wow the blogs posted here are awesome, the octopus and this lama are awesome.
Myself cant seem to get it to work. I think it's not very good at real things. Tried fitness related images, all is weird. Probably with fantasy kinda stuff its better since it has to be less accurate.
NFTs are just numbers on a blockchain. The picture is a canard. In the US I don’t think you can copyright DALL-E images as they aren’t created by a human, so you spend money to make them and anyone else can use them.
I tried a number of these generators a week ago (or so), all with the same prompt: "A child looking longingly at a lollipop on the top shelf" with pretty abysmal (and sometimes horrifying) results. I'm not sure if my expectations are too high, but maybe I was doing it wrong?
>the ball is positioned in such a way that the llama has no real hope of making the shot
I love that we're at the level where the physical "realism" of correctly representing quadrupedals playing basketball is a thing now. I suppose the next level AI will be expected to model a full 3d environment with physical assumptions based on the prompt and then run the simulation
That's the only way to get reliably usable output.
There's a lot of "80% there but not quite" in the current version, which makes it more of a novelty than a useful content generator.
The problem with moving to 3D is there are no almost no 3D data sources that combine textures, poses (where relevant), lighting, 3D geometry and (ideally) physics.
They can be inferred to some extent from 2D sources. But not reliably.
Humans operate effortlessly in 3D and creative humans have no issues with using 3D perceptions creatively.
But as for as most content is concerned it's a 2D world. Which is why AI art bots know the texture of everything and the geometry of nothing.
AI generation is going to be stuck at nearly-but-not-quite until that changes.
While not fully. There is a lot of freely available 3d models that can used as a starting point. Id love a dalle2 for 3d model generation. Even if no texture lighting physics was there.
Boom... Your consciousness is deleted as the DALL-E 4 output for "Evolved monkey person at a computer, wasting time" is delivered to the dinosaur that paid for it.
152 comments
[ 3.0 ms ] story [ 207 ms ] threadI ran out of credits way too fast, so I like to see other people playing with it and their iterative process.
I found this to be the most important point from this piece. Often people don't really know what they really want when it comes to creative work, let alone to some omniscient algorithm. In spite of that, it's a delight to see something you love from an unspecific prompt that you won't find with anything you receive from a human.
Dall.E 2 never ceases to amaze me.
For anyone interested in learning about what Dall.E 2 can do, the author also links to the Dall.E 2 prompt book (discussed in this post https://news.ycombinator.com/item?id=32322329).
[1] https://en.wikipedia.org/wiki/Godwin%27s_law
It needs a better text to image model, I think. Maybe you can fork it and improve?
https://twitter.com/simonw/status/1555626060384911360
https://docs.google.com/presentation/d/1y8EE_p8bw9dIEDguT1bT...
The fact that it's a derivative of an existing work is noteworthy, but I gave it absolutely no guidance on the topic. If i suggest something it will give it a go with similar fervor. eg https://imgur.com/a/N1qWaSV
Here is "llama in a jersey dunking a basketball like Michael Jordan, shot from below, tilted frame, 35°, Dutch angle, extreme long shot, high detail, dramatic backlighting, epic, digital art": https://imgur.com/a/7LoAtRx
Here is "Llama in a jersey dunking a basketball like Michael Jordan, screenshots from the Miyazaki anime movie", much worst: https://imgur.com/a/g99G7Bn
https://parti.research.google/ https://imagen.research.google/
The models themselves are not public however.
Images are more artistic and less clip art-like than Dall-E, but also don’t have a house style like Midjourney. It’s stunningly good - and open source.
What’s really cool is that the devs have worked hard to optimise the model, so after being trained on 1000 A100s it’ll run happily on an 8gb graphics card or M2 Mac.
I had the same trouble. In my experiment I wanted to generate a Porco Rosso style seaplane. illustration. Sadly none of the generated pictured had the whole of the airplane in them. The wingtips or the tail always got left off.
I found this method to be a reliable workaround: I have downloaded the image I liked the most. Used an image editing software to extend the image in the direction I wanted it to be extended and filled the new area with a solid colour. Cropped a 1024x1024 size rectangle such that it had about 40% generated image, and 60% solid colour. Uploaded the new image and asked DALL-E to infill the solid area while leaving the previously generated area unchanged. Selected from the generated extensions the one I liked the best, downloaded it and merged it with the rest of the picture. Repeated the process as required.
You need a generous amount of overlap so the network can figure out which parts is already there and how best to fit the rest. It's a good idea to look at the image segment you need to be infilled. If you as a human can't figure out what it is you are seeing, then the machine won't be able to figure it out either. It will generate something, but it will look out of context once merged.
The other trick I found: I wanted to make my picture a canvas print, and thus I needed a higher resolution image. Higher even then what I can reasonably hope with the above extension trick. What I did is that I have upscaled the image (used bigjpg.com, but there might be better solutions out there.) After that I had a big image, but of course there weren't many small scale details now on it. So I have sliced it up to 1024x1024 rectangles, uploaded the rectangles to DALL-E and asked it to keep the borders intact but redraw the interior of them. This second trick worked particularly well on an area of the picture which shown a city under the airplane. It has added nice small details like windows and doors and roofs with texture without disturbing the overall composition.
What I did:
https://imgur.com/a/QEU7EJ2
With the prefix of the prompt I described the image. I started the extension operations with "red seaplane over fantasy mediterranean city" but then I quickly realised that this was making the network generate floating cities in the sky for me. :D So then I varied the prompt. "red seaplane on blue sky" in the upper regions and "fantasy mediterranean city" in the lower ones.
I went even more specific and used "mediterranean sea port, stone bridge with arches" prefix for a particular detail where I wanted to retain the bridge (which I liked) but improve on the arches. (which looked quite dingy)
(I have just counted and it seems I have used 27 generations for this one project.)
Maybe Dalle-2 is just secretly a studio Ghibli/Miyazaki movie fan.
https://imgur.com/a/U5Hl2gO
I was testing to see how close I could get to replicating a t-shirt graphic concept I saw.
I had been using ~"A telephoto shot of A neglected police car from the 1980s Viewed from a 3/4 angle sits in the distance. The entire vehicle is visible but it is overgrown with grass and flowery vines"
This process sounds great, though it seems like DALLE needs to offer tools to do this automagically.
https://imgur.com/a/YB5StlE
I wonder what Gary Marcus or Filip Pieknewski think about it. Surely they must be eating crow.
What wrote the prompt?
And this AI doesn't. Your anecdote is totally unrelated to the idea of AGI in the gp post. The fact that it made you laugh is a happenstance. It was not "trying" to make you laugh.
I also saw this one recently from Midjourney. Would not call the humor random.
https://www.reddit.com/r/midjourney/comments/w73rhv/prompt_t...
It’s still up on the dalle2 subreddit.
I suspect AGI, depending on how its defined, will be with us in some form in the next few decades at most. Just a hunch. This is nothing to do with that mission though imho. Maybe you can read into it something like, "we are solving lots of discrete problems like this, maybe we can somehow glue them together into a higher level program"? That might give you something AI-esque? My guess is that 'true' AGI will have an elegant solution rather than a big bag of stuff glued together.
An AGI wouldn't need us to this extent, or at all. An AGI would also be able to come up with new ways to represent ideas, even ways that are foreign to us.
We are not. But maybe we are closer to replicating some of our internal brain workings.
I mostly use it and Midjourney for material for my DnD campaign, but I'm going to need to do a little more work to make the whole thing coherent. Only tried it once and it was okay.
The interesting part is that it can do things like "female ice giant" reasonably whereas google will just give you sexy bikini ice giant for stuff like that which is not the vibe of my campaign!
Even when I re-used the exact prompts from the DALL-E Prompt Book, I didn't get anything near the level of quality and fidelity to the prompt that their examples did.
I know it's not a scam, because it's clearly doing amazing stuff under the hood, but I went away thinking that it wasn't as miraculous as it was claimed to be.
There is however one thing to be aware of, the titles posted on /r/dalle2/ and other places are often not the prompts that DALL-E2 got. Instead they are a fun description of the image done by a human after the fact. Random example:
"Chased by an amongus segway"
* https://www.reddit.com/r/dalle2/comments/wkv7za/chased_by_an...
But the actual prompt was:
"Award winning photo of a mole driving a red off road car through a field"
* https://labs.openai.com/s/xnaoxiWeSjiQX1QyVUCHGkl1
Which is quite a bit less impressive, as the actual prompt doesn't really match the image very well. And if you put "Chased by an amongus segway" into DALL-E2, you won't get an image of that quality either.
Here's a result for the prompt of "Woman with green skin, leaves instead of hair, wearing a simple dress, far shot, digital art, hyper-realistic, 8k, ultrahd," for example (all four images)
https://imgur.com/a/f4d8N0u
You will note that none of them are even basically fulfilling the prompt, as well as all four being, in my estimation, ugly and uninteresting. That's not unusual for prompts that involve some element of the fantastic -- though there are corners of less-realistic digital art that it does do well.
https://imgur.com/a/aVbSxHe
You will still note that none of them are far shots, that no depicted character actually has fully green skin, and one of the four has nothing even remotely like leaves for hair. I mean, is it better? Sure. They're less ugly, though none of them are what I'd call great results. But they also aren't really doing a basically competent job of fulfilling the prompt, much less producing a particularly striking or interesting images.
And my point is, outside of a few areas, this is what you get from Dall-E. Lots of misses, and if you're willing to put time into it and work on your results, a few hits. Don't get me wrong, I've gotten stuff from Dall-E that I think is great (I really like this "watercolor painting" for example: https://labs.openai.com/s/AQ7Wy5VHBWcLL5bJ5LbU5SuW) but I think it misrepresents Dall-E to suggest that most of the time it produces basically good images.
I'd say more like, "If you put time and attention into learning its quirks, in its best areas, it'll produce like one in ten images that are basically good."
And, I mean, on some level that's incredible. You can produce 10 images in about three minutes in Dall-E and get some great stuff. But I think people mostly see the top 10% of what Dall-E produces.
It's not fully removing humans from the equation, but you can take something that used to take days and make it a 20 minute operation.
The DALL-E 2 prompt book. If anything, pretty neat look at how the various prompts come out and some of the art created by it.
I've been playing with Stable Diffusion a lot, and in my experience its results are much weaker then what's shown in this post. The artistic pictures that it generates are beautiful, often more beautiful then Dalle-2 ones. But it has a real problem understanding the basic concepts of anything that is not the simplest task like "draw a character in this or that style". And explaining the situations in detail doesn't help - the AI just stumbles upon basic requests.
Seems like Stable Diffusion has a much more shallow understanding of what it draws and can only produce good result for things very similar to the images it learned from. For example, it could generate really good dutch still life paintings for me - with fruits, bottles and all the regular expected objects for this genre of painting. But when I've asked it to add some unusual objects to the painting (like a Nintendo switch, or a laptop) - it couldn't grasp this concept and just added more warbled fruit. Even though the system definitely knows how a Switch looks like.
The results in the post are much more impressive. I doubt that Dalle-2 saw a lot of similar images in training, but in all of the styles and examples it definitely understood how a llama would interact with a basketball, what are their relative sizes and stuff like that. On surface results from different engines might look similar, but to me this is an enormous difference in quality and sophistication.
I think we're at the beginning of exploring what these image models can do and what the best ways to work with them are.
[0] https://promptwiki.com
I thought this was strange. Why hide an AI generated face?
Also, I came across this article which suggests that at some point users were not allowed to share images generating human faces, artificial or not: https://mixed-news.com/en/openais-dall-e-2-may-now-generate-...
That’s not as easy as it sounds. Specially in the surreal cases that DALL-E is usually requested.
Sometimes you don’t know what you want until you see it. Other times you do, but are not able to express in ways that the computer can understand.
I see being able to communicate efficiently with the machine as a future in demand skill
Oddly, the ones it blocked were more sfw than several others it allowed, but of course I don’t know what the outputs would’ve been…
I got blocked a few times with very non sexual prompts, and I suspect that the AI was a bit horny when it interpreted them.
Myself cant seem to get it to work. I think it's not very good at real things. Tried fitness related images, all is weird. Probably with fantasy kinda stuff its better since it has to be less accurate.
I'm more curious of how this will effect stock photography. Soon anyone can generate the exact image they're looking for, no matter how obscure.
I love that we're at the level where the physical "realism" of correctly representing quadrupedals playing basketball is a thing now. I suppose the next level AI will be expected to model a full 3d environment with physical assumptions based on the prompt and then run the simulation
There's a lot of "80% there but not quite" in the current version, which makes it more of a novelty than a useful content generator.
The problem with moving to 3D is there are no almost no 3D data sources that combine textures, poses (where relevant), lighting, 3D geometry and (ideally) physics.
They can be inferred to some extent from 2D sources. But not reliably.
Humans operate effortlessly in 3D and creative humans have no issues with using 3D perceptions creatively.
But as for as most content is concerned it's a 2D world. Which is why AI art bots know the texture of everything and the geometry of nothing.
AI generation is going to be stuck at nearly-but-not-quite until that changes.
Some guy spent hours feeding the AI pictures he liked to get an end result he was happy with.