Finally DALLE-2

October 15, 2022 5:39 PM

DALLE-2 is now open to all. 50 credits free per month. 15 free subsequently.
posted by storybored (53 comments total)

DALLE-2, on the previous FPP:

The social media equivalent of Trump steaks would be a page full of photos of juicy, delicious-looking steaks with the caption "From the Trump Steakhouse – the best steaks in the world!"
posted by snuffleupagus at 5:58 PM on October 15, 2022

prompt book

Contains guides to: aesthetics, vibes, emotional prompt language, photography, film styles, illustration styles, art history, 3D art, prompt engineering, outpainting, inpainting, editing, variations, creating landscape and portrait images in DALL·E, merging images, and more!

posted by aniola at 6:10 PM on October 15, 2022 [3 favorites]

Are they intentionally hiding the image generator? The landing page at beta.openai.com exposes text tools only, so far as I can tell.
posted by snuffleupagus at 6:14 PM on October 15, 2022

Try here
posted by gwint at 6:25 PM on October 15, 2022

I'm signed in; there are no obvious image generation features. [imgur]
posted by snuffleupagus at 6:27 PM on October 15, 2022

Yeah well now I don't want it.
posted by jquinby at 6:39 PM on October 15, 2022 [4 favorites]

Huh. Once you're logged in, go to labs.openai.com, that should do it.
posted by gwint at 6:40 PM on October 15, 2022 [2 favorites]

there we go!

However, the image of the social media equivalent of trump steaks apparently violates the acceptable use policy/SCP prevention interlocks.
posted by snuffleupagus at 6:43 PM on October 15, 2022

If it doesn't let you in because of your location, I've heard that it's pretty easy to bypass this with a proxy or vpn.
posted by signal at 6:45 PM on October 15, 2022

It does require a 'real' phone number; I tried a few VOIP lines and it refused all of them.
posted by snuffleupagus at 6:47 PM on October 15, 2022

However, the image of the social media equivalent of trump steaks apparently violates the acceptable use policy/SCP prevention interlocks.

Invoking SCP here conjured the mental image of "Trump steaks" in a different fashion, less an article of food, more "Eat of this flesh, it is my body". Ew.
posted by notoriety public at 6:52 PM on October 15, 2022 [2 favorites]

I'm making so many paintings of sunflowers and soup.
posted by mittens at 7:09 PM on October 15, 2022 [8 favorites]

I was able to get into DALL-E 2 during the beta stage. For whatever incoherent reason, I balked at paying by the image so instead this month I spent $400 on a video card (RTX 3060 with 12GB memory) and I've been playing with the "invoke ai" fork of Stable Diffusion, which is all open source software & public model. This card is powerful enough to spit out an image every ~15 seconds or so, depending on settings. My previous card, a GTX 960, was about 2 minutes per image, so upgrading ~3 video card generations made a difference.

The last time I updated it, invoke ai gained a really snazzy web ui that I like a lot (not great on mobile though). Apparently I've generated north of 5000 images with it already which would cost $165 so in one sense I'm a third of the way to 'recouping' my 'investment', ignoring the costs of electricity and pretending that 1 SD image ~= 1 DALL-E image for enjoyment value or other measure of usefulness.

Stable Diffusion does great at some things, and poorly at others. (as does DALL-E, though their areas of great, mediocre, and bad differ somewhat). Just earlier when this FPP came out I was trying and failing to get it to do "regular icosahedron"; but then for the purposes of this discussion I also ran that prompt on DALL-E and it can't actually do the job either. SD has no trouble producing images of steaks that range from plausible to "whaaaat is that" but asking it to put that within the form of 'a page full of photos of juicy, delicious-looking steaks with the caption "From Joe's Steakhouse"' is just asking too much; giving the requested style, let alone the caption, just isn't happening. DALL-E does a lot better at the style and at least puts some kind of "Joe-like" lettering in the caption. SD wins on "steaks that look more appetizing on average" though; DALL-E's are overcooked. YMMV.

I find it a lot of fun; you could do worse than working through your credits on DALL-E to find out if you, too, enjoy being able to ask for "huge rooster, chalk drawing, heavy metal album cover" any time of the day or night and have some images a minute later is worth sticking with it after that.
posted by the antecedent of that pronoun at 7:13 PM on October 15, 2022 [8 favorites]

None of the big three - Dall-E 2, Stable Diffusion or Midjourney - accept Paypal, which makes me wonder if Paypal has takes a stance on this whole thing.
posted by BiggerJ at 7:22 PM on October 15, 2022 [1 favorite]

More likely that OpenAI et al don't want to accept whatever shitty terms PayPal imposes. They hardly have to.
posted by snuffleupagus at 7:25 PM on October 15, 2022 [5 favorites]

So far I've tried DALL-E, Midjourney and Stable Diffusion. I seem to get more consistent, better results with Midjourney than DALL-E or Stable Diffusion – but Stable Diffusion can be self-hosted and is great for embiggening the images from Midjourney and occasionally gives better results when taking a first-cut image from Midjourney and then using the img2img tools.

I started playing with Stable Diffusion via the AUTOMATIC1111 web-ui repository on GitHub using a NVIDIA RTX 3080 Ti w/12GB (Windows 10, desktop) and an NVIDIA RTX 2070 Super w/8GB (Linux, laptop). Seems like performance has improved significantly in just the past few weeks. I expect in a year or two self-hosted AI-generated art is going to be very commonplace. Microsoft Designer was just announced this week with DALL-E 2 under the covers.

Note, Stable Diffusion is not strictly speaking "open source." It has use restrictions and other terms in its license that mean it's not conformant to the Open Source Definition. I realize this is a minor distinction to many but it's an important one. I think the usage restrictions are well-meaning but very difficult and unlikely to be enforced.
posted by jzb at 7:27 PM on October 15, 2022 [1 favorite]

Just to note: don't get too attached to anything from dalle2, or remember to download your images. I had signed up earlier and made a bunch of images of Jesus Christ replacing the carburetor in a 1972 Dodge Challenger, a photograph of a deinonychus wearing a cowboy hat relaxing on a chaise longue and reading a trashy romance novel, and several versions of a sea otter wearing a chasuble and celebrating mass in a gothic church.

I clicked on the labs.openai.com link, logged in, and it had forgotten that I existed and set me up with a new account. Which... hey, I get to make more stupid shit for free. But if I'd wanted that keep any of my previous stupid shit I might be miffed.
posted by GCU Sweet and Full of Grace at 7:34 PM on October 15, 2022 [1 favorite]

It is an open beta.

The phone number collection gave me pause, with the VOIP rejection. Why do they care + AI company.
posted by snuffleupagus at 7:36 PM on October 15, 2022 [2 favorites]

Thanks for that clarification, jzb. I know the licensing of the models is complex but naively assumed that the software part (at least right up til you hit proprietary stuff from your GPU vendor) was under a standard license rather than just "on github"
posted by the antecedent of that pronoun at 7:37 PM on October 15, 2022 [1 favorite]

(a simple "the meanest southern baptist preacher in the whole world" reliably produces delightful results)
posted by GCU Sweet and Full of Grace at 7:38 PM on October 15, 2022 [2 favorites]

By the time I got into the beta for DALL-E 2, I was already running Stable Diffusion locally on my M1 Mac. If I can do that, paying for someone else to do the computing doesn't make a lot of sense.

The Diffusion Bee app requires no command line work to install, and it's pretty darn fast on a Mac that doesn't have a dedicated GPU. 57 seconds for a 512x512 image (at 50 steps. It's more than twice as fast as that with the default setting of 25 steps.) The current version I'm running includes 4x upscaling and img2img.
posted by emelenjr at 7:40 PM on October 15, 2022 [3 favorites]

I tried to access DALL-E 2 on the OpenAI site, but from what I can tell it's just the open version of DALL-E 1. At any rate, the results I got were not impressive; much worse on average than what I got from StableDiffusion, let alone NovelAI's image generator. The article above is paywalled, and OpenAI's website makes accessing DALL-E 2 very confusing! Or maybe they're just giving me the run-around because I use a VPN?
posted by ThisIsAThrowaway at 7:47 PM on October 15, 2022

I've played around with Dall-e and MidJourney. I think that I prefer MidJourney. At least it seems to do faces better. Both have issues where you can get 95% of what you want, and the rest of the picture's just wrong.
posted by Spike Glee at 7:55 PM on October 15, 2022 [1 favorite]

I've been having lots of fun with stable diffusion. Low level model access is great for experiments, while 'enter text, get image' feels really limited.
posted by kaibutsu at 8:03 PM on October 15, 2022

Yeah, SD has progressed incredibly fast thanks to the community having access to the whole thing. Dreambooth and hypernetworks are amazing, getting the model to replicate styles and subjects that it doesn't know or didn't even exist when it was initially trained.
posted by simmering octagon at 8:11 PM on October 15, 2022

For the past few months, just for fun, I've been spending more time than I should using DALL-E to make slideshow videos for my brother's songs (1, 2, 3, 4, 5, 6) and accompanying illustrations for a few of my poems (1, 2, 3, 4). To be honest, I'm gobsmacked by how good it is sometimes -- so long as one doesn't try to get anything in particular. It's not unlike working with a certain type of artist: you can tell them in great detail what you want, but in the end of the day, they'll do something totally different that comes from their heart and their vision, so it's good to just be open to enjoying their creativity. That said, I really found DALL-E's inpainting feature to be invaluable. If an image (I usually go for "oil paintings") is really nice in some ways, but has some localized absolutely horrible disaster areas, one can just erase those parts and ask it to try again (sometimes with a modified prompt).

And, although it's a flawed metaphor, when using DALL-E I can't help but ruminate on how much of my own thinking is like half-nonsense "inpainting" from the thing I think I know to the things I imagine might be true.
posted by brambleboy at 9:03 PM on October 15, 2022

I played around with MidJourney until my free credits ran out, generating images for post-apocalyptic RPG Twilight 2000*. They were, after a bit of fiddling, probably good enough to illustrate an RPG book, where you rather want to invoke a mood than have a detailed, coherent illustration of something. I did run into a wall

However, any illustration including vehicles in them fell rather flat, not just that the vehicles it generated had never before been seen on Earth, they often did not very coherent. The prompt "line of abandoned military vehicles rusting in a foggy forest clearing" often ended up with the realization that "no wonder these were abandoned, half of them lack any doors or hatches for entry and the other half have too few wheels to balance themselves". But it was an interesting exercise, and it often nailed the mood of scene.

*) We're playing the 4th edition and it's rather good, if a bit too close for comfort in these times.
posted by Harald74 at 10:11 PM on October 15, 2022 [1 favorite]

I think that's "50 credits free the first month, 15 free subsequently", not "50 credits free per month(?), 15 free subsequently. "
posted by xigxag at 12:08 AM on October 16, 2022

These new digital toys are founded on Mass Art Theft.

Art, Theft.
posted by Faintdreams at 12:58 AM on October 16, 2022 [3 favorites]

Ah, an intellectual-property maximalist?
posted by acb at 5:24 AM on October 16, 2022 [2 favorites]

“These new digital toys are founded on Mass Art Theft.”

Almost all of the tweets on my timeline are this sentiment.

I want to agree and be sympathetic, but I'm not. Sampling in music settled this years ago and, anyway, there's less retained of the original images in the model and its products than are retained with music sampling. The argument is weaker, not stronger. My sense is that many people misunderstand the tech and believe that the source images exist intact in the model, which they very much do not.

This goes to the heart of a lot of intellectual property discussion and these issues are difficult. I disagree with the quoted position, but I don't want to be antagonistic — I think I understand the reasoning and sentiment.

Over the years, I've become more and more skeptical that IP law genuinely protects artists and art. Whether "AI" art is a good or bad thing for art and artists overall, I am unsure. I am inclined to think, though, that wider availability and accessibility has historically been shown to be a big net benefit.
posted by Ivan Fyodorovich at 7:47 AM on October 16, 2022 [5 favorites]

I think evaluating this technology against the law as it's written and interpreted by courts today is missing something important: it should be totally in bounds to say "this is outside anything the law anticipated, it's time to make a new law, what outcomes would we like to see under that hypothetical law".

The world in which an AI startup hoovers up a few TB of data indiscriminately and produces a proprietary model that they charge for access to kinda stinks. They just made something that is (hopefully) hugely valuable, on the backs of millions of individuals, and they aim to profit from it.

It's OK to feel like this is really sus even if you agree with the analysis that under current copyright law the companies that do it are going to go untouched by any and all attempts to show they infringe on someone else's copyright just because of data that was used as an input to the model.
posted by the antecedent of that pronoun at 9:33 AM on October 16, 2022 [2 favorites]

Speaking of sampling in music, is anyone taking this same AI approach to generating samples and loops? Like, I'm curious what kind of music will come out when someone can get samples by querying "funky loop dogs barking" or "theremin ska" or whatever.
posted by RobotHero at 12:05 PM on October 16, 2022

I've also been having fun with Diffusion Bee. Like all of them, I suspect, it hasn't worked out what hands are for or how they work, and if it thinks you want a poster or an album cover it tries to make words all over everything and fails completely, just filling things with letters, which is simultaneously endearing and annoying, like a robot puppy that wants to help you with dinner but only manages to get soup over everything.

I have generated a number of photographs of Victorian Viking Ladies. They're quite intimidating, actually, combining the scarier qualities of Vikings and Victorian Ladies and they carry an impressive array of axe-like weapons. What's interesting is that they're definitely all of those things, and I'm impressed that the application has managed to synthesise the images, for which there's no immediate use, but it's interesting to see them.

At the moment very much at the "raw material for something else" stage.
posted by Grangousier at 12:08 PM on October 16, 2022

These new digital toys are founded on Mass Art Theft.

Art, Theft.


"To steal ideas from one person is plagiarism; to steal from many is research."
posted by Tell Me No Lies at 12:15 PM on October 16, 2022 [5 favorites]

So today was my first contact with this technology in any of its implementations and all I can say is: Holy shit. And technologically it requires quite a lot to astound me. Having studied Machine Learning in depth just makes it more impressive.
posted by Tell Me No Lies at 12:18 PM on October 16, 2022 [3 favorites]

Speaking of sampling in music, is anyone taking this same AI approach to generating samples and loops?

It's very much in its infancy but who knows in a couple of years. I feel it's harder than generating images because if it sounds wrong the whole thing is ruined and annoying to hear, while images can still look right at first glance even if details are wrong.
posted by simmering octagon at 12:18 PM on October 16, 2022

I feel [generating samples and loops] is harder than generating images because if it sounds wrong the whole thing is ruined and annoying to hear, while images can still look right at first glance even if details are wrong.

QFT. The human brain is infinitely more forgiving of visual flaws than audio ones.
posted by Tell Me No Lies at 12:24 PM on October 16, 2022

I wonder if the outcome of Andy Warhol Foundation, Inc. v. Goldsmith will have any bearing on the future of machine generated content.
posted by thedward at 12:33 PM on October 16, 2022 [1 favorite]

Sampling in music settled this years ago
Yes, by insisting that identifiable samples be credited and their creators compensated.

...there's less retained of the original images in the model and its products than are retained with music sampling.
And yet if you ask one of these ML systems for "x in the style of y" it knows what the style of y is, to the point that artists who have developed an individual style are worried they may lose their living.
I'm intensely sceptical of any appeals to the democratizing nature of these tools since the evidence of the last couple of decades shows that software advances end up reducing both quality and choice. As someone on Twitter put it:

"So, I’ll bet very few people in the past who dreamed of a world where AI and robots replaced manual labor so people could work on art never imagined it would be the other way around and we’d be taking double shifts while machines make our art for us."
posted by thatwhichfalls at 12:38 PM on October 16, 2022 [4 favorites]

And yet if you ask one of these ML systems for "x in the style of y" it knows what the style of y is, to the point that artists who have developed an individual style are worried they may lose their living.

But any human who wants to, and who has the appropriate skills, can entirely legally imitate their style right now?
posted by GCU Sweet and Full of Grace at 1:41 PM on October 16, 2022 [2 favorites]

"So, I’ll bet very few people in the past who dreamed of a world where AI and robots replaced manual labor so people could work on art never imagined it would be the other way around and we’d be taking double shifts while machines make our art for us."

I believe it is at this point we are supposed to pull the word "Luddite" out of the bag and swear convincingly that this technology will create new and even more meaningful jobs for artists now that this menial task is largely automated.

That’s the traditional script anyway.
posted by Tell Me No Lies at 1:47 PM on October 16, 2022 [1 favorite]

The outpainting is best use for me so far, not so great at generating specific requested imagery spontaneously. I outpainted from this little render I did of Hunnycub and some background elements. Used Dall-E to start erasing and outpainting to make this big weird landscape I turned into a custom playmat for card games. Here's the image
posted by GoblinHoney at 2:08 PM on October 16, 2022 [1 favorite]

Sampling in music settled this years ago
Yes, by insisting that identifiable samples be credited and their creators compensated.

The key here is identifiable samples. If I chop out 5 seconds of a James Brown track for a song, I owe the rights holder some money. This is more like sampling a huge swath of funk/R&B songs in order to develop tones for a synth. You can copyright a tune, or part of it, but AFAIK you can't copyright the tone of an instrument or style of playing.

If one of these apps spits out an image that contains identifiable chunks of an artists' work, I expect they'd be able to sue. (And I expect that at this early stage that is not out of the question.) But there's not a lot of difference between an artist studying the work of someone they like and then trying to reproduce their style, and Stable Diffusion trying to generate art in someone's style, except speed.

Take a look at this gallery of paintings that are "inspired by" Georgia O'Keeffe. Assuming O'Keeffe were still alive, would you argue that artists trying to reproduce her style shouldn't be allowed to do so?

If a guitarist is trying to sound like David Gilmour and doing a good job of it, assuming they're not actually copying his tunes, should they owe Gilmour royalties?

I've heard a lot of people say that Greta Van Fleet is basically trying to sound like Led Zeppelin but they're writing new songs that Zeppelin never wrote.

What these tools can produce today is neat but no real replacement for an artist, certainly no replacement for someone who wants a specific piece of art or the prestige of something created by a person.

For instance, you might get a passable dragon out of these tools. You aren't going to get a decent illustration with specifics unless you generate a ton of images and do a lot of Photoshop work. If you want "Tiamat the five headed dragon facing down a D&D party consisting of a thief, mage, druid, dwarf fighter and elf in front of a cave with bones and treasure" ... good luck!

This is not to say I'm unsympathetic to artist concerns or that I'm wholly convinced these tools are 100% for the good. Maybe they should restrict "in the style of" prompts to artists in the public domain or at least who are dead.
posted by jzb at 2:29 PM on October 16, 2022 [7 favorites]

I just stopped cold at them wanting my phone number. Absolutely not.
posted by Bottlecap at 3:38 PM on October 16, 2022 [2 favorites]

I just stopped cold at them wanting my phone number. Absolutely not.

If it is a scam it is the most over-the-top honeypot in human history.

I could see not trusting their security though.
posted by Tell Me No Lies at 3:51 PM on October 16, 2022

I think the idea I ran into that sums this stuff up the best is it's not art making, you are commissioning art from a shamelessly derivative creator with a very poor grasp of composition.
I think the comparison to sampling is off base on account of the lack of human skill & decision making involved. at this point: nothing new is being added, all that's coming out is slurry.
posted by velebita at 4:01 PM on October 16, 2022

Mod note: One deleted. Basically "based on this belief system I just made up for you, you and all your countrymen suck and shall surely be destroyed" is ramping it up a bit much? Just speak for yourself, and don't make personal attacks, please.
posted by taz (staff) at 10:04 PM on October 16, 2022 [1 favorite]

But any human who wants to, and who has the appropriate skills, can entirely legally imitate their style right now?

Law exists to constrain the worst of human behavior, I’m not sure it’s the right tool to judge what the best behavior would look like.

I’m with the people who believe this technology is a major milestone, it’s just that I don’t know what it’s a milestone of yet. As always things will unfold however they will, but in this case I’m not sure I have an idea of how they should.
posted by Tell Me No Lies at 6:31 AM on October 17, 2022 [1 favorite]

a very poor grasp of composition

This is not at all the case? I've generated about 16,000 images with Midjourney and my impression is that it generates incredibly compelling compositions. It's really, really good at that part of image-making.
posted by oulipian at 6:49 AM on October 17, 2022

@velebita I'm not sure I agree that it's all "slurry" (some is, for sure) and I think we tend to over-rate human skill as part of art making. That's where the comparison to sampling and so forth come in - all these arguments sound so familiar having heard them already against synths and sampling.

At the time I even agreed until I started hearing music made from samples that really drew me in. (Information Society really turned me around on sampling in music. I remember having a friend/roommate who was a hard-core 70s guitar rock loyalist and hated anything with synths until he roomed with me for a summer and I kept playing Kiss Me, Kiss Me, Kiss Me and Disintegration on my spiffy new 5-CD changer. He finally, grudgingly, asked me to make him a copy of both before he went back to school...)

Here's some things I expect are going to come out of these tools, ultimately:

- It will probably have a major impact on people who turn around quick-hit stuff on platforms like Fiverr and people who do commissions for fans, etc. Not fatal, because the user will still end up having to overcome some hurdle to sign up or install software.

- It's going to make lawyers a lot of money in the interim. I expect people like Greg Rutkowski (maybe not him specifically) will start pressing trademark claims and such. Maybe copyright if they can catch a platform spitting out close copies.

- It's going to have a big impact on stock photos and art once the tools are better and the legality is sorted to the satisfaction of BigCorps. Maybe we'll stop seeing such awful stock art in presentations and sites -- possibly replaced with equally awful AI-generated art, but at least with a little creativity.

I don't see this replacing in-house graphic designers, etc. In any company I've worked at the queue to get work done is always ridiculously long. This may mean more variety and better looking materials. Maybe...
posted by jzb at 8:36 AM on October 17, 2022

Based on the last two days I'm going to need to set a monthly budget cap on this.
posted by Tell Me No Lies at 12:46 PM on October 18, 2022 [1 favorite]

Sooo I just stumbled into the wild world of alternative models for Stable Diffusion, starting from "Waifu Diffusion" and "Lewd Diffusion" and getting more offputting and/or extremely specific from there. No, I don't know how to "try" them. Not linkified form: hxxps://rentry.co/sdmodels
posted by the antecedent of that pronoun at 6:59 PM on October 20, 2022