|
I knew it in my heart of hearts poo poo like this was going to get very good very fast. My argument was always stuff like music or video streaming is going to become this. You pay to have access to some sort of AI generation thing and just..ask it to generate you content on the fly. It's too attractive for a profit-driven company like Disney who hates unions so bad. The issue though is there is only going to be so long its a thing because this is already out of the bag, why Disney AI or Amazon AI when the hobbyist one is even better without weird exclusions? These guys behind this latest Dall-E can try to clamp down on nefarious purposes but they can only protect themselves, this will absolutely be used for all the poo poo you can possibly imagine. The poster that was talking about how things here are going to get very, very weird is also my line of thinking. I can't wait
|
# ¿ Apr 23, 2022 07:36 |
|
|
# ¿ May 14, 2024 16:08 |
|
stratdax posted:If you feed it the same prompt twice in a row, does it return the same result? Like does it decide what the ideal result is, and then that logic doesn't change? Or does it return different results each time because art doesn't have a logical end point. It's always a different result. Might be really similar, might be quite different. There was an example about asking it to show an avacado. Sometimes it was a whole one, sometimes it was sliced in half with the pit visible.
|
# ¿ Apr 24, 2022 22:26 |
|
OpenAi can talk all they want about safeguards and having checks preventing nefarious use but that idea is more to protect them than actually stop any of it. Basically Private Cumshoe posted:what if I tell it to make a naked lady what if it looks like jordan petersons grandma?? is going to be possible very soon as some version of Dall-E-2 will arrive shortly without restrictions.
|
# ¿ May 4, 2022 23:20 |
|
naem posted:it could be argued that art directors will still be making decisions just with this as a time and labor saving tool, there will just be fewer people in the coal mines scribbling away under them Now you're getting to the interesting bit imho. Imagine a shirt that can change what it displays on the fly driven by AI. Hypothetically, how do you enforce copyright on that? I don't think you can, because if AI is involved it's "just" code and has no morals in the game. Whatever you try to program it to do can just as easily be undone, on top of everything being an algorithmic black box you won't even be able to tweak stuff directly, just ~suggest~ things. Ask AI how close you can get before it's copyright infringement and have it only go so far. It's algorithms all the way down here in the future.
|
# ¿ May 5, 2022 18:11 |
|
AI can spit this stuff out so fast though, you don't just get "a" queso mickey, you get one every quarter second.
|
# ¿ May 6, 2022 05:29 |
|
The Alchemist posted:Who cares about copyright in the future, nobody is gonna even watch disney when they can just stay home and go "Computer, give me a cool 2 hour animation of anthropomorphic animals having a wacky adventure saving a princess or something" Yeah it's going to rule
|
# ¿ May 6, 2022 05:58 |
|
LifeSunDeath posted:"russia on fire" lmao @ putin crying
|
# ¿ May 6, 2022 20:01 |
|
|
# ¿ May 8, 2022 08:52 |
|
gently caress You And Diebold posted:Those grandmas looking pretty monochromatic If you're talking about how they are all white then yeah, the ugly side of this AI stuff and the models they are trained on is it, once again, exposes all of our culture's horrible sexist, racist and prejudices! Also why it's cute that OpenAi is like "oh you can't make bad things with this " because the algorithm has the awful poo poo in it, you can't filter out that.
|
# ¿ May 8, 2022 17:34 |
|
Sentient Data posted:The art world will legitimately be a completely different place two years from now, mark my words. This is 100% unironically industrial revolution level poo poo. I'm assuming these models are going to 'break out' as people get more of an understanding on how this all works. Someone mentioned it earlier in the thread that while a bunch of people are looking to capitalize and make money on this very quickly they are going to lose control. Subscriptions to a generative-netflix is one thing, but why not have the model itself run on hardware you own or rent yourself? There is way too much opportunity to undercut everyone at such a fast pace that I'm really curious at how the entertainment industry is going to take this tech getting out even more. This thing can generate fully-realized concept art in seconds. Even just to spit out ideas you have in your head is so goddamn breaking.
|
# ¿ May 13, 2022 05:06 |
|
sigher posted:It's impressive that nerds have found ways to be artistic without needing to be artistic, some of this poo poo is scary good. I don't need to draw anymore, a computer can just poo poo it all out for me. What is needed to take what is in your head and put it out in a way other people can experience how you imagine it is an incredible set of skills, but is that "artistic"? Or is that more of like, learning a language to translate out what you imagine? A discipline? Anything that removes barriers in getting expression out is cool and good. The democratization of art rules.
|
# ¿ May 18, 2022 14:02 |
|
GABA ghoul posted:This paper demonstrates AI that can create 3D environments with textures and correct lighting from just an image. So that you can walk around in your dall-e 2 horror creations with an VR headset. No man's sky 4 is gonna be wild I unironically love this a lot but do feel bad for the straining a lot of industries and professionals that work in them are going to have to deal with.
|
# ¿ May 20, 2022 05:24 |
|
ymgve posted:furry artists are soon out of a job Found this in his feed and cracking up because I'm dumb as hell: peps Peper Pesser PES pepi Peppse Pes Pepen Pess Pessi
|
# ¿ May 24, 2022 02:38 |
|
Inzombiac posted:This tech is growing so quickly that I wonder if in the next 10 years you could enter a prompt and it would spit out a playable (NOT NECESSARILY GOOD) video game. I'm thinking more training an AI to mod games, or training bots to play in particular ways tailored to the human. Seems like when you have a much smaller set of rules in these instances AIs like this can become very good in a short amount of time since you're asking a lot less of it.
|
# ¿ May 29, 2022 07:38 |
|
Granted access to midjourney, the lesser invite-only imagine AI, however it spanks the Dall-E mini. vaporwave baron harkonnen vaporwave baron harkonnen smoking a cigar while riding a surfboard cartoon puppy scared of Pizza Hut effeminate goth johnny bravo A tuna fish's driver's license tuna: dang, i realized tonight at a bar that my drivers license expired haha (this was in a chat I was in, someone named tuna said it) Steamed Hams" VHS tape "World War 95" Microsoft software box 1955 model year DeLorean, technical drawing, blueprint delorean designed by Leonardo da Vinci Leonardo da Vinci brand potato chips Mona Lisa brand running shoes leonard de vinci notebook drawing of curly fries obese self portrait of leonard de vinci, sketch (he spent all his time drawing fries) obese leonard de vinci eating fries, annoyed with the light from the window vaporwave baron harkonnen being fed curly fries by goth johnny bravo (both above the same prompt) Spaceballs the Movie: Anime vaporwave movie poster, title: "GRANDMA II" (both same prompt, worked out)
|
# ¿ Jun 1, 2022 08:02 |
|
cyberpunk beef summer's hottest fashion summer's dumbest fashion Stalin but with long, blonde hair Stalin, but as a woman (I can not get it to generate 'Stalin' without his moustache no matter what I try) Stalin, but anime Bill Clinton as a woman, anime Beyoncé as a man Madonna as a man Big Bird, courtroom sketch, United Kingdom House of Commons cyberpunk nursing home anatomically correct bagel
|
# ¿ Jun 2, 2022 01:30 |
|
See if you can guess the prompt for this one: uber driver sucked me off???
|
# ¿ Jun 2, 2022 22:01 |
|
welcome to ohio we hate you
|
# ¿ Jun 2, 2022 22:23 |
|
Welcome to ________ by Beksinski. See if you can guess the place. First one is a state, next three are cities. Ohio Miami Seattle Chicago
|
# ¿ Jun 2, 2022 23:30 |
|
precision posted:same, exccept i'm gonna make so much money selling custom art to furries The irony I find is these AI systems work on data it's observed, and there is a gently caress-ton of porn, furry and not, already on the internet that you just know is in these models. I was trying various phrases, I tried "told u I was hardcore" and it didn't do it because the word "hardcore". However, you can say 'flashing' and get the thing to show you boobs because that phrase has a specific connotation, same with 'mooning'. Collect them all! Lowtax Fragmaster Genesplicer Huge Gross Burrito Three Olives Grover Grover's House
|
# ¿ Jun 3, 2022 01:30 |
|
TIP posted:drat that's hot try putting "woman" in front of that word
|
# ¿ Jun 3, 2022 01:44 |
|
TIP posted:original was hotter huh, that model is much more modest than the MidJourney one. And when I say boobs I mean approximation of boobs, more like Cronenberg-esque lumpy monsters. Still, you can tell the data is there.
|
# ¿ Jun 3, 2022 01:53 |
|
TIP posted:after a little more testing I determined that there is a simple magic word to make dall-e mini produce breasts lmao MidJourney has a bunch of words flagged, BUT there is a crucial thing they missed: other languages. As far as I can tell it has zero restrictions in Spanish. whoops!
|
# ¿ Jun 3, 2022 03:39 |
|
TO BE CLAR My Lincoln doesn't eat the pizza, he surfs the pizza
|
# ¿ Jun 3, 2022 05:38 |
|
Hyper realistic character concept of twitter dork who is crying, by WLOP 4K, mythical beast concept art, Rendered in Octane, trending in artstation, cgsociety, 4k post-processing highly detailed, Tony Sart, Craig Mullens, dynamic cinematic lighting rendered by octane engine, 8K post processing (copy/pasted other people's prompts, if you load it up with all this garbage it gives it particular looks)
|
# ¿ Jun 3, 2022 05:50 |
|
AI art is cool, good, and my friend. Michael Jackson as Frank Sinatra Lady Gaga as Louis Armstrong John Candy as Ferris Bueller Jack Nicholson as The Terminator Kelsey Grammer as Mr. T Moby as Sasquatch
|
# ¿ Jun 3, 2022 06:58 |
|
TIP posted:got some insights for you A little bundle of joy Happy family life until his parents got divorced when he was 4 Was a good kid in school, no issues, lived with his mom and spent time with is dad on the weekends In high school got into punk, is dad was NOT happy with how he dressed and disowned him. Managed to go off to college, graduated Got married Then went and gone full sovereign citizen after a nasty, nasty divorce after only 1 year
|
# ¿ Jun 3, 2022 18:34 |
|
GABA ghoul posted:Computer, pull the curtains and generate very erotic and sensual pictures of aliens. Computer, dispense lube the future is soon
|
# ¿ Jun 4, 2022 17:20 |
|
https://www.midjourney.com/app/ should take you to your account's gallery which will have the prompts used to generate them (sans the --codes to alter weights and resolutions), handy for when you forget. These 1800s looking diagram ones all share the same prompt, with a single bit changed. Full prompt is: "futuristic ________ design from 1895, mechandrites, steampunk, technical blueprints, line drawing, high detail --aspect 21:9" Place what you want in that space. Xbox desktop PC nuclear submarine flying boat VHS cell phone giant skull Besides that I've been trying to get it to make a picture of a character I have in a PNP game my friends and I play. He's a young adult with one eye. Couldn't get it to understand "one eye" easily, tried "eye patch" which ended up getting a more blended eye-makeup thing more often than not. "pirate eyepatch" however is the key word to generate an eyepatch. Two things have made it abundantly clear when trying to generate good pictures of people: Multiple iterations and the "Light Upscale Redo" which upscales the size of the image without attempting to saturate it with detail. This avoids the problems of the AI trying to shove eyeballs or extra lips where they shouldn't be. Also --aspect 9:16 works well to tag on the end of your prompt to get a more portrait-style aspect that allows the AI to draw more of the character. Prompts for these: hyper realistic character concept of young adult medieval peasant, long hair, pirate eyepatch, fantasy, by WLOP 4K, mythical beast concept art, Rendered in Octane, cgsociety, highly detailed, Tony Sart, Craig Mullens, dynamic cinematic lighting, --aspect 9:16 KakerMix fucked around with this message at 18:50 on Jun 4, 2022 |
# ¿ Jun 4, 2022 18:27 |
|
Sekenr posted:Can someone get me a "photorealistic emperor Paul Muad'dib posing gloriously" This is the stuff that clearly points to the different models as the AI I'm using clearly knows 'Muad'dib' since that's the only place it will pull the desert stuff from, which then stands to reason there is probably a lot of Dune art its been taught on. Here is MidJourney's very first attempt at this prompt: I chose the second one and upscaled it: The eyes aren't quite right, take this upscaled version and put it through "Light Upscale Redo" Better. It's like the model de-HDRs it making it honestly look nicer. More detail = not always better as clearly demonstrated. Those are portraits though, so gently caress with the prompt some more and go over a bunch of iterations and you can get something probably more toward what you'd actually be after: Here is some other Dune-specific prompts which shows what data it actually has, first attempts only and just the bit in quotes: "wormsign" "stillsuit" "Bene Gesserit" "Ornithopter" "Sandworm" "Fremen"
|
# ¿ Jun 5, 2022 17:48 |
|
The Fattest PI posted:Oh man I hope I get accepted into the real Dall-E-2, the dall-e mini isn't enough. This poo poo is insane Maybe all these AIs aren't quite awake yet
|
# ¿ Jun 6, 2022 17:27 |
|
Muppets! Putin Stalin Trump Boris Johnson The Weeknd
|
# ¿ Jun 6, 2022 22:08 |
|
Carl Sagan as Elon Musk Elon Musk homeless Barack and Biden as Mao Zedong Barack and Biden as Kim II-sung Bernie Sanders as Lenin Mao Zedong blazing it Bernie Sanders as an 80s glam rock god ymgve posted:not good at context MidJourney is a ton better than dall-e mini, though I like that dall-e mini seems to handle certain things better anyway plus as far as I can tell, no text restrictions. I can't do anything with "Al Gore" because "gore" is filtered >:| sunrise on mars
|
# ¿ Jun 9, 2022 02:52 |
|
Taco Bell velvet elvis Elvis as an 80s glam rock god A Strange Aeon posted:Can someone do a still from an unreleased live action Calvin and Hobbes movie by a director of your choice? Denis Villeneuve David Lowery Robert Eggers Pixar KakerMix fucked around with this message at 05:13 on Jun 9, 2022 |
# ¿ Jun 9, 2022 05:08 |
|
ymgve posted:I've been trying to make a "just download and run" package for Dall-E Mini - I haven't tested in on a fresh Windows install, but it should in theory only need the Cuda 11.7 toolkit, then just run the .bat file: When I get home I'll give this a shot, thanks a ton for doing this.
|
# ¿ Jun 9, 2022 20:54 |
|
ymgve posted:Great! Could you tell me your hardware and your performance? It takes about 10 seconds to generate an image on a 3080ti here ~12 seconds for me on a 3090
|
# ¿ Jun 10, 2022 09:00 |
|
spleen merchant posted:Could someone please do "a 1920's heavy metal concert on the moon"? Hot Pink Justice posted:It really wants to put the moon in the background, it also is coming up with a lot of album art: I changed the prompt to "a 1920 heavy metal concert on the lunar surface" which made it so it didn't always try putting the moon in the sky, since you'd be on it.
|
# ¿ Jun 10, 2022 23:16 |
|
"Columbo arresting Jessica Fletcher for her crimes" "cybernetic dad telling a joke to bored kids" "portrait of a "Dana Scully" in the style of Basquiat and hannah höch" "Arrakis, a fusion of art deco and star wars architecture, artstation, volumetric lighting, by james clyne::4 ralph mcquarrie::1, realistic beautiful high detail render" (swiped this prompt from someone else and added 'Arrakis') "futuristic 70s Soviet Union cigarette ad with pretty woman smoking and cool catch phrase" "futuristic, 70s West German cigarette ad with pretty woman smoking and cool catch phrase" "stylized poster, general Sherman, vaporwave" "stylized poster of futuristic 80s Soviet Union soft drink" "designed by Syd Mead" series: robot lawyer cellphone energy drink computer mainframe laser rifle hightop sneaker cybernetic dog cybernetic dad
|
# ¿ Jun 13, 2022 07:23 |
|
Dang It Bhabhi! posted:how come yours are better There are three major AI art things right now. Dall-E mini, which is the one you see all over social media and in this thread with the hosed up faces. Then there is Dall-E 2, the one with the news articles and bears doing trigonometry and all that. If you see someone saying 'Dall-E' and it looks choice, it's the invite-only Dall-E 2. I'm using the third wheel sorta one, MidJourney, that is also invite-only but not as strong or capable as Dall-E 2. It's extremely strong with mixing ideas, doing portraits, and creating whole new art in the styles of whatever you ask it to, like Syd Mead. Not so much with doing goofy context stuff like even Dall-E mini is decent at.
|
# ¿ Jun 14, 2022 02:34 |
|
|
# ¿ May 14, 2024 16:08 |
|
Tunicate posted:What kind of prompt babble (trending on artstation, etc) are you adding for midjourney? If you're in the discord (discord being how you interact with MidJourney at all) I'll just see what other people are generating in the various channels and swipe their prompts. It's a nice way to see what the AI does with certain prompts. "strong" artists like Syd Mead or Zdzisław Beksiński are so strong that you can get away with saying "thing by Zdzisław Beksiński" and it will kick out stellar results every time. You can do the same thing with anime or 'rendered in octane' to get the AI to attempt to output something in a realistic style. There is a ton of weighting you can do with the options and things to really tweak what the AI will spit out. The other part is I'm cherry picking things, Syd Mead stuff is usually good right from the start, but most things require perhaps 3-4 passes to get something that works, then maybe 3-4 more passes on that specific out put. For every good picture you see there might be 20~ others you don't see that didn't look right. TIP posted:There's also Disco Diffusion which you can run in a Google colab at that link. Yeah this one as well, I messed with it briefly before I got my invite. Let's see what MidJourney does with these exact same prompts. I'll run it till I get something presentable and will tell how many revisions it took to get there. quote:"A beautiful wilderness in which a lonely monolith glowing, by Asher Brown Durand" First output out of the four you get at once. Looking at Asher Brown Durand's art I'd say that Midjourney gets the artistic style better, though both of the AIs "get" the artist since they both output essentially the same composition. Here is the first output, this one was #1: quote:"A scenic view of the pleasure dome of Kubla Khan in the tropical jungle of Xanadu, by Asher Brown Durand" Similar situation here, but with the dome itself being from a poem so there are different interpretations of it. Picked #4 out of the first output since I feel it matches closer to the one done by Disco Diffusion. Here is the first output for this one: quote:"A close-up view of a man meeting the devil, by Caravaggio" Again MidJourney seems to get really close to the artist's style. Took a bit to get this prompt to spit out something I'd be happy with. Here are the first two: Poor results, imho. I added "face to face" to the prompt and got this: and I went with #4 which is the upsized one you see above. quote:"A lady in a long white dress lying amongst many large snakes, tropical jungle, by Prateep Kochabua" Context like this is where MidJourney just isn't very good and gets showed up by even Dall-E mini. The AI INSISTS on making the woman in the dress a part of the snakes and the snakes a part of the woman. Everything looked like this, even when I changed the prompt around trying to get it to generate something decent. Changed the prompt to "woman in white dress sleeping on a bed of snakes, tropical jungle, by Prateep Kochabua" and got better results, but still bad: 11 passes with the original and variations of the original prompt with nothing to show for it. Changed the prompt to "woman with snake, jungle, by Prateep Kochabua" and got better results, but not what Disco was able to pull off. 7 prompts with a couple dead-ends. MidJourney still insisted on blending her arm with the snake, just completely unable to keep it all separated. Like this: The one from Disco isn't 100% correct, but those snakes are not the woman, nor is she the snakes. I gave up.
|
# ¿ Jun 14, 2022 04:32 |