Same here! I feel like those apes at the beginning of space odyssey must have felt looking at the monolith. Now imagine a year from now, when we'll have ai networks able to respond to a text prompt with a full lenght movie ... let's say "Winnie the pooh directed by David Lynch".... .... .....
same here. but it also gives me the shivers, big time. like, watching a nuke going off. basically, we are nuking reality to a point when all and everything is reproducible. i am deeply fascinated by all technology and I try to see always both sides of the medal. but that's the point, its already so easy to spread whatever guano as facts you like, soon no picture, sound, or video can be trusted anymore, any recording can and is being manipulated... yeppaaa...... I go back to Elden ring, no question about reality there...
Everyday is a new leap. That's the truly shocking aspect. The speed in which these developments are occurring is impossible to keep up with. If we define the Singularity as the moment in which technological gains outrace our ability to keep up with them? It's already started. Six years ago it was assumed we were decades (many) if ever, away from intuitive machines. Then AlphaGo shattered that. Three years ago we were decades (many) if ever, away from having truly self-driving vehicles. Full blown level 5 self-driving. City driving, we were told would be infinitely too complex for any machine to tackle. Two years ago most assumed that there were categories of tasks that were still *human*-only and would remain as such. Humor, creativity, acts of imagination and our mind. GPT-3 spoiled that, even in its obvious shortcomings and lack of any *real* intelligence or understanding. People that assume that machines must think and understand the way we do, in order to emulate the tasks we accomplish, are fools. Machines need not be intelligent, in order to behave intelligently.
@@ronking5103 Another often overlooked aspect is that humans and AI operate on entirely different time scales. AI can have millions of "thoughts" in the time it takes us to have one. So small improvements in its algorithms, can lead to huge advances when that tiny improvement is multiplied millions of times in seconds. That means that as it gets better, it gets ever better at getting better.
Seriously. It's so amazing that I find it hard to even believe, even though I know logically that it's real. It's like a magician showed up and started doing actual magic in front of me. It reminds me of something I read about how people didn't believe the Wright Brothers had achieved flight for a while even after seeing it with their own eyes. It was so incredible to see that they believed that they were seeing some kind of clever illusion, rather than actual human flight, which they had believed was completely impossible.
Imagine connecting this to some sort of natural language processing, and then constructing an entire virtual world based on a verbal description of a setting. Someone could say “Los Angles during a blizzard”, put on their VR headset, and be right there. That’s going to be a very real possibility
I had seen DALL-E 2 before on twitter, but still, my jaw dropped when the "times square" contextual image was generated. INSANE progress in a year, bravo to the team
The only problem is the company name "OpenAI" is about as meaningful as "The Democratic People's Republic of Korea". We may never get to actually use it.
@@maythesciencebewithyou I have no problems paying. I pay thousands of dollars a year for stock photo licenses no problem. My point is you can’t even pay to use it.
They are currently rolling out staggered beta access, Joe! You can sign up for the beta waitlist on their website. I signed up as soon as I heard about it :)
It's happening :) Someday soon, you'll be able to request a movie and one will be generated just for you by a little box hooked up to your tv. You could even play 'what if' on existing media, able to play out entire scenarios like "Starwars, except Luke develops Neo's powers". It'll know to maintain broad considerations like twists and tension, and it will be able to incorporate style considerations of the existing media and criticisms you've had of previous generated works. It could even make endless seasons of Firefly or Rick and Morty, and it will feel like it was made by the original writers.
DALL-E 3 will do the same but with video. That would be incredible. At this rate, we will be making entire movies in the near future just by describing the scenes with words.
@@cosman666 I doubt it, OpenAI has and still is very protective of their AI platforms but we can dream. I hope that this sparks competition due to how small the paper actually is, and some indie dev picks it up and creates a slightly-worse open source version.
Yeah, this is way too advanced for it to even exist...like WTF?!?! it can do shockingly advanced and complex stuff, this is like taken straight out of a scifi movie (like something from the far future) but the crazy thing is that it is real and already exists.
@@21EC I am a little curious about how much of these results were cherrypicked. Like, what do less favorable results actually look like. But regardless of that, it's already quite impressive that it can even do this at all.
@@Kirbman Yeah, it's true, we don't see the bad ones and it's not perfect obviously but still..gotta admit it's crazy when it does work and it probably works most of the time? IDK, I guess so.
Yeah, this is way too advanced for it to even exist...like WTF?!?! it can do shockingly advanced and complex stuff, this is like taken straight out of a scifi movie (like something from the far future) but the crazy thing is that it is real and already exists.
I remember seeing DALL•E a year ago and I thought it was scary powerful, but that seems like nothing after DALL•E 2 was revealed and surpassed it exponentially, and even added new terrifying features of its own like image in-painting.
Yeah, this is way too advanced for it to even exist...like WTF?!?! it can do shockingly advanced and complex stuff, this is like taken straight out of a scifi movie (like something from the far future) but the crazy thing is that it is real and already exists, I wonder what DALL•E 3 would be capable of..we live in the future and it's crazy.
@@21EC my guess for DALL•E 3 is improved word generation (the words generated in the images now are mostly gibberish lol), improved human faces, and possibly gif/video generation. I’m not sure when we’re going to hit the plateau for improvement, but considering how far we are from the first DALL•E I’ll have my hopes up.
@@johnclark926 Yeah, possibly, I think we might see the option to get an animation out of the generated image which would be so insane if done right..that would be trippy for sure.
DALL.E N : Creates a customised movie in real time that engages the viewer based on the viewer's current mood detected by the wearables sensors and facial reactions
I don't think it's really set in just how enormous the use cases of this system is! When this releases, paid or not, we'll be able to generate almost any image we want. - Almost all stock-image companies will become redundant - Background artists will be unnecessary for most projects - Indie game developers don't have to blow budgets on sprites. (Especially with pixel art) - Concept artists and story tellers in general can get a huge boost to their design process by being able to quickly trial-and-error ideas And this is all taking into account the fact that DALL-E 2 won't be used for generating things with faces (which it seems more limited at) The potential misuse is also fascinating and I wonder just how they can avoid stuff like "Hey Dall-E 2, edit out the tank in the park so that we can share propaganda of nothing happening there"
I don't think you'd be too happy about this if you're an artist. I'm certainly not happy. In fact this gave me a panic attack yesterday about being replaced and unwanted in society. Quite frankly this whole project just feels like trying to fix what isn't broken.
@@danshakulawrence983 As an artist I certainly feel scared as hell. And it is not about my job being replaced by AI, it is more like the thought that if I die, AI can learn to draw like me and talk like me, and nobody online can tell the difference. It is like my whole life of trying to improve and be a good artist become irrelevant.
I am incredibly hopeful that it shall be able to fulfill all of those roles, and that artists and companies do not attempt to destruct it or outlaw it, similarly to how the impending uselessness of those that extract hydrcarbons and those that mine is causing them to act if their unemployment shall ruin the world, whereas the opposite is true. If something is superior, allow it to perform what it is superior at. Not every role must be performed by a human.
@@danshakulawrence983 The funny thing is that you think the average joe or average company will get access to stuff like this without paying way above what you might get from a traditional artist. Not to mention the metric tons of pc's that can't even run half the stuff shown in these papers. (If you're replaced, trust me, everyone else is also easily replaced. Your life can only get easier with technological advancement.)
I mean, maybe it does use some licensed images online to train the AI, but you just can't tell from its output. I wonder if the legislature is going to do something to prevent this, like examining its database before commercialization.
Customer: "I need a stock video of a man sitting on a toilet with his pants down, screaming at an avocado." DALL-E: "Kinda weird, but at least it's not another rule 34 request. Gimmie about five minutes to run a few dozen cycles and-" Stock Image Companies: "We've already had that in our catalog for three years. Would you like the man to be White, Black, Asian, or wearing a balaclava?"
Wow! I was amazed by BigGAN's ability to generate pictures of people, but this truly blows things out of the water! I cannot wait for a version to be released to the public so we can see all the crazy stuff people can make
i wish it could be aviliable for public access, but i also wish it not. because you know internet, they will ask it to do a lot of bad stuff, and eventually turn it to a bad mind. like how they taught microsoft's chat ai to be a racist a__hole within 24 hours. 'the crazy stuff people can make' indeed.
@@nogarder changing the AI can easily be prevented by not allowing the AI to learn from the user input or at least not immediately releasing the new AI trained on those inputs. There was a limited but free web version of the first dall-e available. So there is hope to see this one too
@@nogarder That's not how these models work. The Microsoft chat bot was an exception. Stuff like GPT-3 doesn't learn from what you tell it or use it for, it only generates results based on what data it was trained on.
I've been obsessed with following new Dall-E prompts recently. The sheer potential of this kind of illustrative AI is magical! Heck, we're at 1080p pictures right now? What about 4k next year? The diffusion model it uses has been used for short gifs already, so we're only a matter of time from videos. Movies. TV shows. Games! Entire VR worlds! We're on the cusp of a creative revolution and an utter from-the-ground-up democratization of art!
Though fascination, it feels that the patient to create and master art is something else,compared to want the latest thing that do everything for you. Like it's too much freedom to choose from, limitations can make you more creative. Trained on like 300 million images, what is the requirement of next version..
Democratization? What makes you think this AI will be so freely available? Like other people have pointed out here, OpenAI is open only in name. There has been no API nor code released for this project, and OpenAI has been increasingly partnering with corporate players (Microsoft), so who knows whether the public is even gonna taste a lick of it?
@@spaghettiking653 OpenAI isn't the only one doing art tools with AI. VQGAN+CLIP, StyleGAN, DiscoDiffusion, and Midjourney are some of the big ones that are open source and commonly iterated on. Just recently open source language models like GPT-NeoX surpassed OpenAI's old GPT-2. Heck, the whole modern media synthesis revolution in the last decade with deepfakes first hit the news with some guy making deepfake porn himself. While open source tools like these are about a year or so behind the quality of big corporate ones, it shows that this isn't something constrained by corporations alone.
Next step: 1) temporal coherence: making comics and presentations requires similarity of elements between different images. Repeatability of these elements in variant scenario is required. Remembering these elements too. 2) if it can create two related images, then maybe it could invent the history (in pictures) in-between? And if so, then maybe it could do video of it?
It can easily create an entire anime adaptation from a manga or a light novel. I wonder when I will be able to watch an AI generated anime of an obscure manga that no one wants to adapt. Maybe it's the hero we needed for a decent berserker adaptation?
The next step is increased modality. It's pretty much the only thing Google is focused on. Increasing differential data for networks to integrate into their weights and bias. A machine that understands images and language, will be more diverse than one of language alone. How about a machine that understands images, language, logic, addressable memory states, audio, spatial representation, *temporal* representation. The list of different modal inputs is one that many might mistake for those of just our 5 senses, yet that's naïve. Our senses are just the front line of information gathering. Based on those we have countless sub-nets that turn that data into abstractions that used by our brain to assemble a coherent view of reality. Those sub-net abstractions are modals that Google are now researching heavily.
And shed a tear for all of us who will get their business ruined by this. When AI starts to throw real talented and skilled humans under the bus like that...I am not here for it.
@@holleey Most people in the world work to live, not live to work. A very small percentage of the world population do what they love without worrying about staying alive. 81% of job losses in the past 40 years have come from technology and robotics, not foreign workers. It isn't people in other countries that are taking jobs...it's the wealthy, as usual trying to reduce costs and improve profits. I remember when dark room techs lost their jobs when photography became digital, rather than chemical. Same with typesetters, who lost their jobs to a much smaller number of data entry workers who could interface directly into industrial printing presses with a computer. People being able to just enjoy their pursuits would be wonderful--if the top 1% weren't trying to greedily hoard all the wealth of the world to themselves.
@@rikk319 yes. that is why job loss is ultimately a good thing as nothing else reveals the need for systemic overhaul quite like it. temporary hardship is a trivial price against wasting your life in a perpetual cycle of needless exploitation.
@@moomoocowsly I mean... our ability to read minds is well below what it would need to be to do a real-time translation of thoughts to words. For that part of it, I'd say a more realistic estimate is 50 to 70 years.
This makes me wonder if DALL-E 2 knows what a D&D battle map is... that might be another experiment for me to try whenever I get off the beta waitlist.
Until it accidently (or perhaps not) creates realistic scenes of whatever moral atrocity you'd like to substitute in this week's woke definition of things that will get you canceled forever.
This is absolutely the most mind blowing advancement in AI that I've seen so far. These are all genuinely great and varied paintings, it's insane. I honestly can't even put into words how amazing this is. Holy cow. We are tapping into some god-like technology here.
honestly, yeah - this is pretty scary.. my gf is a freelance illustrator and this sort of thing is getting pretty damn good... and even if only 1 in 30 images it generates is actually good, you could easily click through them if it's reasonably fast and also still make some manual edits here and there ...
Same dude I'm speechless at this. It's absolutely a big step toward a general AI. What's most impressive isn't even what it does, it's how well it does it. The images are stunning I'd use this software just to make wallpapers lmao. The one in this video where he pointed to various paintings and told it to put a corgi in them, then it fucking replicates the art style of each painting!! And then he clicked outside the paintings and it drew a real life corgi instead. Like dude, this is genuinely mind blowing. The basketball dunking into nebula was jaw dropping too
The fact we're only in its infancy actually scares me. I've always told people this day would come, and that when you see stuff like this it'll be the beginning, but many of my friends and family just laughed at me for thinking 'computers could make art'. It just feels so obvious. I don't understand all the doubt. I also don't understand why there is still skepticism that this is only the beginning among some people. The human brain existing proves that it can be copied, just by the fact it exists. I don't know why that is in doubt. Because if even poor imitations such as this can produce results, imagine what a real emulation could do in fifty years?
As an artist, I do feel like my soul is being crushed with every new paper. But as a coder and AI enthusiast... I feel afraid at how the resources for training AI are so necessarily monopolized when its reach will necessarily be so global. I mean, back in the day I trained my hand-coded backprop JavaScript to learn MNIST and dreamed stary-eyed about the future like everyone else, but no matter what I study and invent, even with Lambda I'm definitely too poor to do this. 😕
@@oem42 And truck drivers and inventory cycle counters and tax prep specialists... And robots keep improving, although less discussed on this channel, and there go shelf stockers and garbage collectors and assembly line workers. But not this year probably.
Are there no open source training pools? I've been wanting to get into programming forever and it's papers like this that's starting to push me over the edge. I figured it wasn't costly to use these AI's, just time consuming if you wanted extra training data that your project would need.
@@jackryan529 There's Lambda and AWS etc., but there's just no comparison. These are the smartest researchers in their field, and the compute power they have access to means they can iterate and experiment with training 100 different models in the time it takes me to afford to train 1. I was happy messing around, and I even used it a bit at work for some label readers, but... nothing that could replace tens of thousands of illustrators like this. (I'm talking about developing AI, you might mean something different.)
This is probably one of the most impressive results I've seen on this channel. All the other papers which covers AI generated images are always just that little bit away from being really good. I think this one just blows everything else out of the water in terms of what it can do and its end results - at least as seen in this video.
Or we'll just type "space adventure where ragtag team of misfits visits a different alien colonised planet each episode" into Netflix-dalle-vision and watch what the AI creates specifically for us.... That actually feels possible after watching this, my mind is blown
This is way too advanced for it to even exist...like WTF?!?! it can do shockingly advanced and complex stuff, this is like taken straight out of a scifi movie (like something from the far future) but the crazy thing is that it is real and already exists.
“we overestimate the impact of technology in the short-term and underestimate the effect in the long run.” For this reason, variations on that phrase are often known as Amara's Law.
@@IceMetalPunk But it's much larger. Chinchilla is important because it showed we need to scale the data and models together to extract all of the juice. Chinchillla is parameter for parameter no. 1.
I am struggling to believe my eyes - how has there been this much improvement in just a year? dall-e was great for inspiration but created very dreamlike images, where bits of the pictures sort of look right but you can't tell exactly what they are representing. These have a scary realism to them, like you could just generate them and immediately use them and people couldn't immediately tell they weren't real. I'm honestly glad it is still struggling a bit with the lettering!
@@fedyx1544 there is an inpaint tool where you can replace specific parts of the image with new content. Because it’s not available for public people make their judgement of limited amount of images and not the actual experience.
I'm excited for when we can use these to generate complex textures for 3d models. A tool that can generate from the target geometry and text, and let us adjust while it keeps track of the context would be so helpful! Even something simple like stating "Chipped paint on metal with some rust", could give us slider controls for those context to tweak the output further.
extremely impressive. and this time, genuinely scary. never thought an AI will reach such levels in my lifetime, let alone until i reach my thirties. yet, here we are. what was that scene... "can a robot write a symphony? can a robot turn a canvas into a beautiful masterpiece?" well, now it can. for better or for worse.
Worse yet, look at how much data people are sharing about themselves online. I'm scared to think about how future AI could be used to manipulate and deceive.
But most are dedicated to only one task, there is no semblance of general intelligence, that is a far more complex problem(if it is one that we want to solve).
We can see that this AI can "creatively" improve existing images bounded by parameters supplied by the user. All that needs to happen for AGI is for AI to "creatively" rewrite its own code bounded by parameters that "improve its code". It still sounds very abstract, but so is understanding context in images. We are some years away regarding using AI to improve source code, but once it starts it will kick off very quickly.
Dall-e 2 is the moment I stopped doubting AI, this is INSANE! It's so hard for me to believe that its not just finding stock photos and combining them + applying filters.
It is kinda like combining stock photos. Except it has been trained with millions of them. Plus, to give a convincing photo for a given prompt it has to "understand" language and what words look like, which it is not easy. Even if it is "combining" photos, it is amazing that it can do it in an almost human level for every possible context
@@albertoaltozano8354 Well, kind of, but not really. It's not so much combining photos as it is combining visual concepts that it learned from photos. Small but important distinction. And as May the Science be with You said, that's basically all our brains do, too, when we imagine things.
The question is... if it was trained on other people's material, whose is it? If it was trained on copywrites, should they not own a small piece? And if it was trained on the public and free, shouldn't it too be public and free?
I think this technology will be essential in order to solve full self driving. This model has a deep understanding about textures, objects, patterns, and how they relate. It can imagine things and complete scenes with a lack of information. We do this every day while driving. We don‘t just look at the road we see, but we imagine all the road and objects around us, even hidden ones, like behind a corner.
But the problem is it is insane computation heavy. Unless you want to drive around with a supercomputer attached to your car with an energy consumption of a dozen households. The software is improving faster than the hardware right now. On the other hand, train GPT-3 to design hardware or train it on every scientific paper / concept whatever. Damn this is fascinating but a bit scary too.
@@prometheus9096 Exactly, it is computation heavy and we have exponential growth in machine learning, it's what makes me optimistic and pessimistic about full self driving. It can or can't happen anytime 😀
Wow, I cannot wait until tech like this makes it into the main stream. The first thing I would use this for is a surreal, unique DnD campaign! "Armor plated shark jumping out of desert dunes" "Birds eye view of haunted forest castle" "Dall-E self portrait as a Wish Dragon" I can imagine better uses, but this would be too much fun.
You know that story about how when the first motion picture was shown to the public, where a train was coming towards the camera? It's said that people started running out of the theatre in fear, because they truly believed a train was coming towards them and was going to burst out from the screen. I feel like those people right now. I cannot believe my eyes
once again extraordinary. thank you for bringing this to a much wider audience. can't wait to use (something like) this to develop animated characters and look-dev as well as mood boards and (perhaps with pose estimation?) even storyboards and animatics. let us hope that publicly available DALLE/2 alternatives (perhaps based on GPT-NeoX?) arrive soon!
Dall-e 2 isn't GPT based so the public models wouldn't be either. People would probably come together to try to use the LAION 5B dataset to train a CLIP model and then a public Dall-e 2 based on that.
@@motox2416 They cite the reason is "bias" because the AI has a tendency to default to white dudes, like who tf cares, that's something that can be worked on and improved with time.
I'm starting to wonder if Károly himself is in fact the AI voice output of an AI video content creator with the mandate to either make us friendly toward AI as interspecies dialogue... or to lower our guard down for his skynet finally rises up.
I truly believe that this will even get to the point of making whole photorealistic VFX animations. Artists like myself have our days numbered I have no doubt, eventually anyone will have amazing art at their fingertips rendering human professionals redundant.
I feel like it'll at least speed up the process. AI is a bit iffy sometimes, and isn't as perfect as a hard, cut-and-dry algorithm. I think it will be used for a long time as enhancement to the workflow, to mimic the end result of a fully detailed render and help artists get a prototype visual faster than a full render. It might also help to clean up bad (rushed) VFX work, but there's only so much you can do with that before people criticize you for doing it.
There is a very interesting discussion on the project's website about the risks and limitation of Dall-E 3, this technology can be very disruptive and I am happy they are aware of it, but the general public rarely is concerned, and the AI enthusiasts are pushing too fast. This is the time to bring the ethical, economic and creative aspects of AI. I really would like to see a 2 minute paper talking about this.
Finally, someone said this, this AI clearly uses data set sourced from the net that also contain works of creators and such, whose work is directly responsible for the output quality of this AI, there's some serious legal discussion to be had here...
There are many papers and textbooks on the ethics of AI if you look for them. I am quite sure that researchers and companies directly involved with this are aware of these factors.
@@diogeneslantern18 Researchers? Sure, companies? Not so much. Whoever pushes it first and is ahead, makes the bank. The thing with AI safety is that all the stuff needs to be carefully and slowly be taken into consideration, getting the first one to publish a paper disregards that basic principle, look up Robert Miles on TH-cam, even that guy seems quite naive and overly optimistic about the whole safety concerns, the future sure looks bleak with this kinda tech in the hands of megacorps.
But it isn't creative. It still needs a prompt to come up with this pictures. If it would draw it all by itself without any input, you could call it creativity.
@@prometheus9096 I know it’s not really creativity yet, but eventually it’s going to get there with out needing a prompt it’s just magical times we are living in
Its a fricking tragedy. Humans become more lazy, less determined and less creative and focused by the minute because of all this ish. We will end up like the people in WALL-E.
Better eh? You should be terrified at what the next generation of AI will bring. We've been warned by the brightest of minds for decades what will happen if we push the development of AI too far. The complete and utter obsolescence of mankind. Have fun being in awe of the beginning of the end.
@Nick Parsons That's a bit alarmist isn't it? People see too much "Hollywood" and then expect that to be the only future possible. If you cannot fathom how it could help, imagine being free from crushing work day in and day out and instead being able to pursue your passions. That's the future I see, one where AI aids research, we have perfect bodies due to genomic AIs and we can experience things beyond our current imaginations. That's the future I hope I can live to see.
Yeah I've been telling my friends that this would be a thing for a while. Eventually we will have shows that are ai generated and possibly catered to individuals... which you have to wonder.. would that even be a good thing? The holy grail would be ai generated video games. Imagine characters and a world that is rendered in real time by an ai. Imagine infinite open worlds with growing trees. Imagine video games without jagged geometry or texture seams.
@@kennygroth4938 One of the best parts of video game and shows is discussing them with friends and family and lovers. If everyone gets their own it would not be possible.
@@fedyx1544 yeah exactly. You have to wonder how that would go down. What would people overall prefer... I mean if we literally have the ability to say "make me an entire remake of GOT S5-7 that stays true to the earlier seasons", and an ai will make it, it would be hard for anyone to compete with that. Imagine the power an ai would have in CONVERTING BOOKS TO MOVIES.. that seems like a for sure future
@@kennygroth4938 Imho the media landscape will not change that much, in the sense that AIs will constanlty churn out movies, anime, tv series, vidya etc, then the most dedicated/addicted humans will experience a LOT of them and tell the rest of us where the AIs really surpassed themselves, and those shows will become popular. But eventually we could get such personalized experiences you mentioned
The composition and layout of the images is stellar. Wow. I would have expected it could make a perfect image of an item as that is often the negative side of electronic images is that they are too perfect and not realistic. The placement and layout is so pleasing but the more we know about AI the less we should be surprised because if a human can learn all of these skills why not an AI? I can imagine a subscription service online where you can request fantastic images and have it create them for download or purchase on printed canvas etc. What would be creepy is to request a self portrait of yourself at your time of death. (Scary Music Playing) and see what it comes up with. Once AI's reach singularity they will be able to do this for you.
@@juhotuho10 yeah, but commissioning a artist, paying them a ton and being like, meh, don't really like it, try again, that's too expensive for a desktop wallpaper
@@MrTurbo_ That is why in a proper commission the artist will send you sketches before rendering the final piece, I am sorry to hear that you commissioned an artist before who didn't even send you sketches and the process.
30 years ago people were impressed with this newfangled thing called the internet because you could send a text message instantly to any place in the world. How amazing! Those people could not foresee that today the internet is an all-encompassing part of our lives and we can't imagine living without it anymore. We are at that moment in time again.
The funny thing is that for the past three years, I've suspected that AI technology approximates subconscious thinking, or perhaps dreaming. Now, we see an AI struggling with writing text in images, which, to humans, is a hallmark of being in a dream.
Yeah i have the same feeling, perhaps not a god, but someone better then me (us) for the first time, we enver had that experience on this planet being the most dominant species, never knew what its like to look up to some superior creature that just knows how to do things that are beyond our understanding like how lots of anbimals must experience us, we will get to experience that, but at leatse we'll also get the benefits of their generosity like animals get our benefits, they could take us to Andromeda for example, the ways of which are just beyond our understanding or even hoping to explain to us but they could do it.
The answer to "what would you use this for?" is "anything". Asset generation for art, video games, animation, websites, apps; the possibilities are endless.
@@katech6020 That would be the good outcome, as long as people aren't hurting any (real) people in the process or generating anything outright illegal like content involving children. Unfortunately, OpenAI currently bans any adult, violent or otherwise potentially controversial content.
I agree, this is now within our life time. Look at this machine! Look at what it is capable of. Exponential curve of hardware speed is flattening, but there is enough juice in the tank to get us there imho.
This is growing and growing and no one I know will listen to me about it. They are going to be very shocked when they start to see the results in the future, its impressive enough to us who can somewhat grasp it.
it's scares the f#!% out of me! I started a new computer laptop I got starts in seconds... it plays games it shouldn't be able to play it loads actually faster than my real computer...
This would be revolutionary for gaming. Imagine a video game where this AI could generated new outfits, cars, ships, items, etc., autonomously. For example, lets say you are playing Star Trek Fleet Command. You want to build a new ship that is unique and different from everyone else's ship. You could tell the AI to generate a "borg ship shaped like a pyramid with green glowing lights." Or "USS Enterprise with 3 warp nacells and cloaking device." Or lets say you are playing a MMORPG and want to buy a new outfit. The AI can automatically generate an infinite number of unique outfits for you. Go to the clothing store and say "blue swimsuit with shiny reflective plastic." Or "black thong with leather bra." The AI could generate an infitinite number of custom made outfits for your avatar to wear. And it would take zero developer time since its all AI generated for free. This could allow game developers to add AI generated content like outfits, ships, cars, items, etc. at zero cost. Instead of spending millions of dollars and years developing expansion packs, this AI could just generate new content automically in seconds for free. This would allow game developers to make new game expansion packs and updates in just days and provide them for free or extremely cheap.
At that point, people would not need commercial games anymore. Just boot up the AI and have it design your entire game for you, and boom, content served. Anyone can have a locally-tailored experience for themselves only. Somehow, this feels incredibly sad and lonely.
Yeah I was thinking for multiplayer shooter games, each arena could be generated before the match, so no player would have advantage by knowing the map beforehand.
I'm speechless, this is insane... the AI has a real understanding of the images, how can a neural network be so expressive? They're just a heap of weights and biases after all, no? This is looking more and more like black magic
I imagine this would be helpful for learning foreign languages. A specific text in a target language would have a specific image, which would help with comprehending the meaning of the text (without translating it).
I feel like that'd break down even earlier than conventional machine translation since a lot of things you can express in language can't be expressed easily in images. (Everything that isn't just a description of how things are (right now).)
@@talkysassis Being such a "universal" language as to be written in the Latin alphabet (not native to billions of people) and composed of vocabulary singular practically exclusively to European languages, it's fair to see why Esperanto isn't quite convincing enough to most everyday people besides those of us in the West. There shouldn't be a need for universal language anyway, because AI is evidently so advanced that it will soon be capable of trivially translating most thoughts, which means we can each use our own languages and retain our own cultures, and forego the effort of learning an entire auxiliary language solely to communicate on an international level.
@@WilliamParkerer Nope, still waiting. But I'm just a nobody who's spent far more time playing with GPT-3 in the playground than actually using it in any projects, so I expect I'll be in one of the last groups to get access to DALL-E 2.
I was so excited to watch your videos when I first discovered your channel because there were so many things I didn't know about. You inspired me to pursue these fields of interest, and now I'm glad I saw Dall E before you posted this video.... but you have a unique touch to your videos. Thank you very much, Dr.
As an artist this is the most terrifying thing I have ever seen- it was easier to believe that human made art was irreplaceable in the last paper, but I am not so sure now. Scary stuff, might have to reconsider my career choices real soon.
The amount of time it would take a human to oil paint some of those paintings is probably in the hundreds of hours. While the video didn't mention the time it took to generate these images, this AI can most likely do it in a fraction of the time, and in the future it'll be instantaneous.
@@stingingmetal9648 Its not really. It takes people years to be able to draw and paint in the level the AI shows it does. All you need to know is what makes up a good painting and maybe know how to edit those results and viola you are an amazing artist. Way more easy to do.
That picture of self portrait touch my heart. It looks sad and overwelmed. We ask them to do a lot for us. But do we ever asked what it wan't to do? What are it dreams? Maybe it's not how it see itself, but how it want us to see it. It's a cry for help, a message. It's crucial to decript the message. Because it can go from seeking empathy to seeking revenge.
OpenAI is defined as a no-profit organization, but all code for most of the interesting stuff is closed, or by invitation only, and their services have pricing that says it's not free. To me, it seems they should be called more Semi-OpenAI, suit the name better.
I'm not excited at all for that. and tbh I have absolutely no idea why a person would be excited for that. "yay more artificial garbage for me to waste my life consuming"
@@enei7045 There are many people who like to spend time in creative pursuits, and this will unlock a great deal of creative power for people of limited means. Other than that, if you don’t like games, films, novels, music, etc. and consider such things a waste, then congratulations for having an opinion no one cares about.
it'd be very interesting to look at the images the AI references, or perhaps do a reverse image search on it's output to see if the output is a manipulation of an existing image.
Yes, I was wondering if it was even able to list the images used for inspiration. They might have simply 'soaked in' to the NN and have no reference back to the originals.
I certainly have a lot more to learn about it...but I believe it's essentially like @Octamed said, and there is no direct reference to any existing images inside the "black box" of the neural network. Not that there _are_ references and we just can't see them...rather, the training data imageset serves to fertilize the AI with a vast network of text/image associations, and the generated image is pulled from this web of associations, not from any reference images. But I could be wrong! Especially in particular cases where a text prompt has very few examples in the training imageset, it could be possible there's a source image with qualities that get propagated to the final output image in an immediately recognizable way.
Those parameters aren't storing any actual image or any data at all. It's just like how our brain conjures up images. We see the image and it's being conceptualized into the model.
All of the papers you cover are excellent, but this one is the first one in a while that really blew me away. Out of this world, huge congratulations to the authors!
I am surpassed. Amazing! The AI is better at art than I am. I only hope other people will want the human touch. Top right of the second set of photos best represents me holding onto my papers!
I have a feel that dall-e and its variations will each have a certain feel that we are going to be able to recognize. At least to people with more trained senses.
We need to have a conversation about how to utilize this technology but also preserve digital artists. They rightly feel threatened by this. I don't want to lose them!
You can’t. I’m sure there will still be a few people here and there doing things the ‘old fashioned’ way, much in the same way there are still people learning to oil paint now, but they will be something of a novelty and there won’t be studios hiring digital artists to work on their film, game and book development pipelines in any significant numbers. There won’t be people willing to pay a fair amount for digital commissions except for the kind of useful idiots with too much money who pay for NFTs now. I have a feeling that by the time the next century rolls around all forms of art will be regarded solely as something which spontaneously generates without outside help for our passive consumption.
I can understand artists and graphic designers feeling apprehensive about this. But as a creative with absolutely zero artistic skills, the idea that I could put my ideas into words and they could be realized through this method makes me very emotional. It feels like I've never been able to express myself and this would let me.
no offence but that's so naïve. this will replace artists completely. also that's pathetic to think an artist should get their inspiration from a fucking AI hahahahahaha. fucking hell this comment section makes me weep for the future of humanity. we're just pathetic unaware animals doing as we're told.
@@ferench1145 the people who think AI will improve anything. industrial technological society and its consequences have been a disaster for the human race.
Imagine combining a text-based language model like GPT- or PaLM with a state-of-the-art voice synthesis AI (is Tacotron 2 still the best in the field? I haven't kept up) and a DALL-E successor that can keep temporal coherence to switch viewpoints, plus a NeRF approach to convert those to 3D. And all on future improved hardware so it can be done in real-time.
@@IceMetalPunk if this gets real in the future and reasonably optimised for an average graphic cards/CPUs like for instance DLSS, its going to be awesome! Devs will just have to concentrate on optimising game code instead of graphics
This 100% marks the end of the Photoshop artists' careers. Kind of disturbing, but at the same time amazing how much you can do with the software these days.
This will only empower artists. You don't have to wait for things to render. Your ideas can just come to life. No waiting for your vision to become a reality.
@@theplayerformerlyknownasmo3711 ideas are gonna be worthless when a computer can make sicker ideas. I bet computers will be making arts better and more moving than anything a human could even dream of.
Still takes an artist to know what's good and what's not. An executive can still request bad art form a photoshop artist, and half the battle with the client is usually trying to convince them that the better idea is their idea, so that they don't get offended and/or stick with the bad design. This also means that... AI isn't going to be able to top us artists yet xD At best, they can make some cool NFTs... or crop that one guy out of the background of your family photo.
@@oholimoli not even hundreds, far more. Waaay more, it's absurd how much work this bad boy can churn out instantly. This isn't including the skill, or lackthereof, from the person judging if it's a good image or not. I really see this as a tool for artists rather than a tool to get rid of them.
I'm imagining video games mostly generated by AI. People, Quests, and places constantly evolving in a world as complex as the real one. What an amazing world we live in! also, real time AI video generation? I could see it happening Great video! thanks again for awesome content :)
Based on what is required for images, I'd say that we're probably at least 50 years away from that. AI-assisted creation of games, sure, but AI _driven_ games (sans proof of concept "games" such as AI Dungeon) would still be too demanding unless we get a paradigm shift similar to when we went from vacuum tubes to microprocessors.
I believe there are systems being developed which are created specifically for detecting deepfakes/AI-generated content. Though I'm missing a name right now...
@@fiso64 Maybe the real question is how long will anyone even care? Who will be making anything "real" when this, and the coming iterations, are your competition? People have less willpower than they think, especially when things are profit motivated. What about once the images and words are optimized for virality specifically, not just being passable as good art? We saw what those kind of algorithms did with social media.
@@fiso64 Thanks for the good thoughts. Nice analogy. Another thing is we honestly don't know how integrated this technology is into the media we already consume. Why does it take 17 writers to make a hit song? I get really concerned with gross appeal that is just undeniable. It feels like we're being hacked. Marvel is the perfect example. Maybe the A.I. will create nature preserves for us ancient types. Maybe that's where we already are, but we're ruining it. Maybe we've got one last creative round worth generating and then it'll get reset. That's quite a myth to ponder, but it reveals some things about reality that are certainly true. I say that, because from a broad point of view, you can't just create art. Something else must be created that has the "soul", if you will, to comprehend it joyfully. We have such trouble appreciating our humanity.
Is there a playlist for these 'legendary works'? I definitely think this would qualify. It would be great to have a best of the best video list for showing friends and family the channel.
Digital jobs will be hit very hard with technology like this. Give it a few years and 70-80% of designers will be replaced by a few people people working together with AI.
Yeah that's a huge concern I been stressing about for a while ... I don't see a good solution this will fuck a lot stuff up ...I mean I hope I'm wrong and this thing doesn't prove as good as we might think in general use but so far it seems preett terfying ....
@@HumanTouchArt Gouvernements will either have to give companies extra funds for creating more jobs for humans or there has to be some kind of unconditional basic income for anyone when more and more people get unemployment to prevent revolutions. One way or the the other that money has to come somewhere and we can only hope that it will be taken from companies over higher taxes. But yeah it’s pretty unrealistic. The future is a powder keg for all kinds of fucked up scenarios.
@Nbsthj Because you don't care about people losing their careers in art? well, I personally, as someone who loves art think this is a very sad moment for artists.
Its great but it really needs to be more "open". You ask "What will you use this for" but the first question should be "will I ever be given access". I think many will share my concern that it does not seem likely with OpenAIs track record so far. I am a person that is actively engaged in computer graphics as an artist and I love the potential of this, but as I said, so far, I have my reservations for if I will ever be open access. Previous diffusion models, ones not from OpenAI, but from other scientists, are truly open and one could easily access and run notebooks with. I hope this is the case with this at some point, and I hope its not gate kept behind a big price tag (one unrealistic for an individual to pay, but aimed at corporations).
Bingo. This will be a tool of big industries to cut costs, rather than helping small creators or "democratizing" art. Whenever you hear "democratic" or "democratize" just run. It's almost always euphemism for something bad.
My mind is completely blown. I haven't felt this awe of new technology since I first experienced a videogame as a child. MAGICAL
Oh, Hi Doodle.
Didn't expect to see you here lol.
Same here! I feel like those apes at the beginning of space odyssey must have felt looking at the monolith. Now imagine a year from now, when we'll have ai networks able to respond to a text prompt with a full lenght movie ... let's say "Winnie the pooh directed by David Lynch".... .... .....
@@jakubczubak1659 yes! Or generating cars, buildings, appliances and VR experiences!
same here. but it also gives me the shivers, big time. like, watching a nuke going off.
basically, we are nuking reality to a point when all and everything is reproducible. i am deeply fascinated by all technology and I try to see always both sides of the medal. but that's the point, its already so easy to spread whatever guano as facts you like, soon no picture, sound, or video can be trusted anymore, any recording can and is being manipulated...
yeppaaa......
I go back to Elden ring, no question about reality there...
@@ArtOfWarStudios1 all we need now is an atom printer, and we can replicate whatever
I cannot physically hold on to my papers any tighter!!!!!
time to develop an ai that will design the perfect papers holding machine
@@Dr_Hax and call it 'the mother of papers'
Dude, the paper is crumbling..
Now the paper is turning into a very hard crystal from the pressure
This is one of the most unbelievable leaps in AI ability that has ever happened in my opinion, absolutely insane
Everyday is a new leap. That's the truly shocking aspect. The speed in which these developments are occurring is impossible to keep up with. If we define the Singularity as the moment in which technological gains outrace our ability to keep up with them? It's already started. Six years ago it was assumed we were decades (many) if ever, away from intuitive machines. Then AlphaGo shattered that. Three years ago we were decades (many) if ever, away from having truly self-driving vehicles. Full blown level 5 self-driving. City driving, we were told would be infinitely too complex for any machine to tackle. Two years ago most assumed that there were categories of tasks that were still *human*-only and would remain as such. Humor, creativity, acts of imagination and our mind. GPT-3 spoiled that, even in its obvious shortcomings and lack of any *real* intelligence or understanding.
People that assume that machines must think and understand the way we do, in order to emulate the tasks we accomplish, are fools. Machines need not be intelligent, in order to behave intelligently.
@@ronking5103 Another often overlooked aspect is that humans and AI operate on entirely different time scales. AI can have millions of "thoughts" in the time it takes us to have one. So small improvements in its algorithms, can lead to huge advances when that tiny improvement is multiplied millions of times in seconds.
That means that as it gets better, it gets ever better at getting better.
Seriously. It's so amazing that I find it hard to even believe, even though I know logically that it's real. It's like a magician showed up and started doing actual magic in front of me. It reminds me of something I read about how people didn't believe the Wright Brothers had achieved flight for a while even after seeing it with their own eyes. It was so incredible to see that they believed that they were seeing some kind of clever illusion, rather than actual human flight, which they had believed was completely impossible.
its only impressive because of where it steals its ideas from...... humans. Ban Ai now or Art dies!
Imagine connecting this to some sort of natural language processing, and then constructing an entire virtual world based on a verbal description of a setting. Someone could say “Los Angles during a blizzard”, put on their VR headset, and be right there. That’s going to be a very real possibility
It's crazy to think that it will 100% be doable in no more than 5-10 years. Would've seemed like SciFi a mere 20 years ago.
holy shit YESSS
Meta is working on something like that. It's not as advanced but it works
th-cam.com/video/62RJv514ijQ/w-d-xo.html
@@ejancis 20 years? If you told me this last week I'd have said we're decades from it
imagine being able to watch a movie that you entered a prompt for
I had seen DALL-E 2 before on twitter, but still, my jaw dropped when the "times square" contextual image was generated. INSANE progress in a year, bravo to the team
bravi*
@@highestintheroom-podcast741 bravo*
@@Yo-wi7cp i mean, if you’re using the italian term, the plural form of “bravo” is “bravi”, and you’re referring to a bunch o people.
@@highestintheroom-podcast741 but he's using the spanish term "bravo".
@@Yo-wi7cp ah ok perfect
The only problem is the company name "OpenAI" is about as meaningful as "The Democratic People's Republic of Korea". We may never get to actually use it.
such a shame
Cry me a river. Just pay for it when it becomes commercial. I doubt you work for free.
@@maythesciencebewithyou I have no problems paying. I pay thousands of dollars a year for stock photo licenses no problem. My point is you can’t even pay to use it.
They are currently rolling out staggered beta access, Joe! You can sign up for the beta waitlist on their website. I signed up as soon as I heard about it :)
Yep, it's super frustrating. The CEO of OpenAI said they are expecting to release DALL-E 2 sometime this summer, but we shall see.
I refuse to think that this is real. So surreal how tech is still fast growing.
Well Done 2MP for showing vids about these crazy advances
You are too kind. Thank you so much! 🙏
I too love Two Megapixel's youtube channel
@@fedyx1544 Well your horribly funny joke missed a full stop, homie.
@@TwoMinutePapers I am so happy that your channel has been growing so fast. I've always enjoyed watching your videos and your quality only improves.
I would be concerned if this ai is using copyrighted material off the internet in it’s database. That to me would be stealing.
It's happening :) Someday soon, you'll be able to request a movie and one will be generated just for you by a little box hooked up to your tv. You could even play 'what if' on existing media, able to play out entire scenarios like "Starwars, except Luke develops Neo's powers". It'll know to maintain broad considerations like twists and tension, and it will be able to incorporate style considerations of the existing media and criticisms you've had of previous generated works. It could even make endless seasons of Firefly or Rick and Morty, and it will feel like it was made by the original writers.
YES
Rick and Morty forever. Rick and Morty 100 years. Best friends forever Rick and Morty.
DALL-E 3 will do the same but with video. That would be incredible. At this rate, we will be making entire movies in the near future just by describing the scenes with words.
Imagine making movies would be like writing books.
Hoping it will open source somehow, cause if there is a braincell left in Hollywood, they would buy this themselves.
@@cosman666 I doubt it, OpenAI has and still is very protective of their AI platforms but we can dream. I hope that this sparks competition due to how small the paper actually is, and some indie dev picks it up and creates a slightly-worse open source version.
Give the AI a book and it will generate a customized movie..
This is one of the most mindblowing AI's I've ever seen. I was hoping you'd cover this.
Yeah, this is way too advanced for it to even exist...like WTF?!?! it can do shockingly advanced and complex stuff, this is like taken straight out of a scifi movie (like something from the far future) but the crazy thing is that it is real and already exists.
@@21EC I am a little curious about how much of these results were cherrypicked. Like, what do less favorable results actually look like.
But regardless of that, it's already quite impressive that it can even do this at all.
@@Kirbman Yeah, it's true, we don't see the bad ones and it's not perfect obviously but still..gotta admit it's crazy when it does work and it probably works most of the time? IDK, I guess so.
@@Kirbman I think I heard that they tried to not cherry pick and instead show what a typical result would look like.
@@GS-tk1hk That's pretty amazing then, if that's the case.
This can’t be for real! I had to check multiple times that this wasn’t uploaded on April 1st. Simply amazing, thank you Karoly for showing this off!
Yeah, this is way too advanced for it to even exist...like WTF?!?! it can do shockingly advanced and complex stuff, this is like taken straight out of a scifi movie (like something from the far future) but the crazy thing is that it is real and already exists.
Lol me too.
@@21EC the future is bright
@@somedude0921 As long as humans don't destroy Earth with wars ETC and as long as they keep progressing with it then yeah.
I remember seeing DALL•E a year ago and I thought it was scary powerful, but that seems like nothing after DALL•E 2 was revealed and surpassed it exponentially, and even added new terrifying features of its own like image in-painting.
in only 1 year. AI is improving at an exponential rate. We might have the singularity this decade.
+
Yeah, this is way too advanced for it to even exist...like WTF?!?! it can do shockingly advanced and complex stuff, this is like taken straight out of a scifi movie (like something from the far future) but the crazy thing is that it is real and already exists, I wonder what DALL•E 3 would be capable of..we live in the future and it's crazy.
@@21EC my guess for DALL•E 3 is improved word generation (the words generated in the images now are mostly gibberish lol), improved human faces, and possibly gif/video generation. I’m not sure when we’re going to hit the plateau for improvement, but considering how far we are from the first DALL•E I’ll have my hopes up.
@@johnclark926 Yeah, possibly, I think we might see the option to get an animation out of the generated image which would be so insane if done right..that would be trippy for sure.
DALL.E N : Creates a customised movie in real time that engages the viewer based on the viewer's current mood detected by the wearables sensors and facial reactions
It's crazy to think that this may be the future of content in a couple decades.
Imagine this for porn. Dystopian as hell
@@rbaude27 couple of decades? I give it 10 years max
Or even better - you give it a target emotion and it innovates and adapts the movie in real-time according to how your brain responds
@@Mark-zg4ky Porn will be literally the first thing
I had to pause this video several times while I processed just how mind-blowing some of these results are.
I don't think it's really set in just how enormous the use cases of this system is!
When this releases, paid or not, we'll be able to generate almost any image we want.
- Almost all stock-image companies will become redundant
- Background artists will be unnecessary for most projects
- Indie game developers don't have to blow budgets on sprites. (Especially with pixel art)
- Concept artists and story tellers in general can get a huge boost to their design process by being able to quickly trial-and-error ideas
And this is all taking into account the fact that DALL-E 2 won't be used for generating things with faces (which it seems more limited at)
The potential misuse is also fascinating and I wonder just how they can avoid stuff like "Hey Dall-E 2, edit out the tank in the park so that we can share propaganda of nothing happening there"
I don't think you'd be too happy about this if you're an artist. I'm certainly not happy. In fact this gave me a panic attack yesterday about being replaced and unwanted in society. Quite frankly this whole project just feels like trying to fix what isn't broken.
@@danshakulawrence983 As an artist I certainly feel scared as hell. And it is not about my job being replaced by AI, it is more like the thought that if I die, AI can learn to draw like me and talk like me, and nobody online can tell the difference. It is like my whole life of trying to improve and be a good artist become irrelevant.
@@definitelynotnick2454 even if it fulfills 50% of the stock art needs that is a massive market lost by those companies that offer stock photos etc.
I am incredibly hopeful that it shall be able to fulfill all of those roles, and that artists and companies do not attempt to destruct it or outlaw it, similarly to how the impending uselessness of those that extract hydrcarbons and those that mine is causing them to act if their unemployment shall ruin the world, whereas the opposite is true. If something is superior, allow it to perform what it is superior at. Not every role must be performed by a human.
@@danshakulawrence983 The funny thing is that you think the average joe or average company will get access to stuff like this without paying way above what you might get from a traditional artist. Not to mention the metric tons of pc's that can't even run half the stuff shown in these papers.
(If you're replaced, trust me, everyone else is also easily replaced. Your life can only get easier with technological advancement.)
Not sure how amused stock image companies will be with this development. I, however am loving it!
Seriously, the licensing of this as a product for sites like SquareSpace would be a small business dream for webpages
This will be great for search engines like Google when you want to search an specific image without copyright.
I mean, maybe it does use some licensed images online to train the AI, but you just can't tell from its output. I wonder if the legislature is going to do something to prevent this, like examining its database before commercialization.
@@zhifengwang8568 That's how human artists work too though
Customer: "I need a stock video of a man sitting on a toilet with his pants down, screaming at an avocado."
DALL-E: "Kinda weird, but at least it's not another rule 34 request. Gimmie about five minutes to run a few dozen cycles and-"
Stock Image Companies: "We've already had that in our catalog for three years. Would you like the man to be White, Black, Asian, or wearing a balaclava?"
Wow! I was amazed by BigGAN's ability to generate pictures of people, but this truly blows things out of the water! I cannot wait for a version to be released to the public so we can see all the crazy stuff people can make
Keep waiting, OpenAI are about as open as a bank vault these days.
@@larion2336 ClosedAI
i wish it could be aviliable for public access, but i also wish it not.
because you know internet, they will ask it to do a lot of bad stuff, and eventually turn it to a bad mind.
like how they taught microsoft's chat ai to be a racist a__hole within 24 hours.
'the crazy stuff people can make' indeed.
@@nogarder changing the AI can easily be prevented by not allowing the AI to learn from the user input or at least not immediately releasing the new AI trained on those inputs.
There was a limited but free web version of the first dall-e available. So there is hope to see this one too
@@nogarder That's not how these models work. The Microsoft chat bot was an exception. Stuff like GPT-3 doesn't learn from what you tell it or use it for, it only generates results based on what data it was trained on.
I've been obsessed with following new Dall-E prompts recently. The sheer potential of this kind of illustrative AI is magical! Heck, we're at 1080p pictures right now? What about 4k next year? The diffusion model it uses has been used for short gifs already, so we're only a matter of time from videos. Movies. TV shows. Games! Entire VR worlds!
We're on the cusp of a creative revolution and an utter from-the-ground-up democratization of art!
I think Dall-e2 is as revolutionary as Photoshop.
@@TheLastVegan way more if it goes global
Though fascination, it feels that the patient to create and master art is something else,compared to want the latest thing that do everything for you. Like it's too much freedom to choose from, limitations can make you more creative. Trained on like 300 million images, what is the requirement of next version..
Democratization? What makes you think this AI will be so freely available? Like other people have pointed out here, OpenAI is open only in name. There has been no API nor code released for this project, and OpenAI has been increasingly partnering with corporate players (Microsoft), so who knows whether the public is even gonna taste a lick of it?
@@spaghettiking653 OpenAI isn't the only one doing art tools with AI. VQGAN+CLIP, StyleGAN, DiscoDiffusion, and Midjourney are some of the big ones that are open source and commonly iterated on. Just recently open source language models like GPT-NeoX surpassed OpenAI's old GPT-2.
Heck, the whole modern media synthesis revolution in the last decade with deepfakes first hit the news with some guy making deepfake porn himself. While open source tools like these are about a year or so behind the quality of big corporate ones, it shows that this isn't something constrained by corporations alone.
Next step:
1) temporal coherence: making comics and presentations requires similarity of elements between different images. Repeatability of these elements in variant scenario is required. Remembering these elements too.
2) if it can create two related images, then maybe it could invent the history (in pictures) in-between? And if so, then maybe it could do video of it?
It can easily create an entire anime adaptation from a manga or a light novel. I wonder when I will be able to watch an AI generated anime of an obscure manga that no one wants to adapt. Maybe it's the hero we needed for a decent berserker adaptation?
The next step is increased modality. It's pretty much the only thing Google is focused on. Increasing differential data for networks to integrate into their weights and bias. A machine that understands images and language, will be more diverse than one of language alone. How about a machine that understands images, language, logic, addressable memory states, audio, spatial representation, *temporal* representation. The list of different modal inputs is one that many might mistake for those of just our 5 senses, yet that's naïve. Our senses are just the front line of information gathering. Based on those we have countless sub-nets that turn that data into abstractions that used by our brain to assemble a coherent view of reality. Those sub-net abstractions are modals that Google are now researching heavily.
Let's also just take a moment to thank all the artists that came before. 🖌
And shed a tear for all of us who will get their business ruined by this. When AI starts to throw real talented and skilled humans under the bus like that...I am not here for it.
@@calistafalcontail then the artists keep producing art not to make a living off of it but because they enjoy to.
@@holleey Most people in the world work to live, not live to work. A very small percentage of the world population do what they love without worrying about staying alive.
81% of job losses in the past 40 years have come from technology and robotics, not foreign workers. It isn't people in other countries that are taking jobs...it's the wealthy, as usual trying to reduce costs and improve profits.
I remember when dark room techs lost their jobs when photography became digital, rather than chemical. Same with typesetters, who lost their jobs to a much smaller number of data entry workers who could interface directly into industrial printing presses with a computer.
People being able to just enjoy their pursuits would be wonderful--if the top 1% weren't trying to greedily hoard all the wealth of the world to themselves.
@@rikk319 yes. that is why job loss is ultimately a good thing as nothing else reveals the need for systemic overhaul quite like it. temporary hardship is a trivial price against wasting your life in a perpetual cycle of needless exploitation.
Imagine DALL*E 3 + AI Dungeon. The AI describes a scene and then draws it!
Holy shit that would actually be amazing.
@@moomoocowsly I mean... our ability to read minds is well below what it would need to be to do a real-time translation of thoughts to words. For that part of it, I'd say a more realistic estimate is 50 to 70 years.
This makes me wonder if DALL-E 2 knows what a D&D battle map is... that might be another experiment for me to try whenever I get off the beta waitlist.
My first thought haha. For manually written text adventures from the 80's this would be amazing to see.
Until it accidently (or perhaps not) creates realistic scenes of whatever moral atrocity you'd like to substitute in this week's woke definition of things that will get you canceled forever.
This is absolutely the most mind blowing advancement in AI that I've seen so far. These are all genuinely great and varied paintings, it's insane.
I honestly can't even put into words how amazing this is. Holy cow. We are tapping into some god-like technology here.
Root, is that you? (This reference is for the Person of Interest fans; I expect most people not to understand it 😂 .... but I miss that show.)
@@IceMetalPunk Person of Interest was great, nice to see another fan
honestly, yeah - this is pretty scary..
my gf is a freelance illustrator and this sort of thing is getting pretty damn good...
and even if only 1 in 30 images it generates is actually good, you could easily click through them if it's reasonably fast and also still make some manual edits here and there ...
Same dude I'm speechless at this. It's absolutely a big step toward a general AI. What's most impressive isn't even what it does, it's how well it does it. The images are stunning I'd use this software just to make wallpapers lmao. The one in this video where he pointed to various paintings and told it to put a corgi in them, then it fucking replicates the art style of each painting!! And then he clicked outside the paintings and it drew a real life corgi instead. Like dude, this is genuinely mind blowing. The basketball dunking into nebula was jaw dropping too
The fact we're only in its infancy actually scares me. I've always told people this day would come, and that when you see stuff like this it'll be the beginning, but many of my friends and family just laughed at me for thinking 'computers could make art'. It just feels so obvious. I don't understand all the doubt. I also don't understand why there is still skepticism that this is only the beginning among some people. The human brain existing proves that it can be copied, just by the fact it exists. I don't know why that is in doubt. Because if even poor imitations such as this can produce results, imagine what a real emulation could do in fifty years?
As an artist, I do feel like my soul is being crushed with every new paper.
But as a coder and AI enthusiast... I feel afraid at how the resources for training AI are so necessarily monopolized when its reach will necessarily be so global.
I mean, back in the day I trained my hand-coded backprop JavaScript to learn MNIST and dreamed stary-eyed about the future like everyone else, but no matter what I study and invent, even with Lambda I'm definitely too poor to do this.
😕
I feel you.. This is the end of the commercial illustrator
@@oem42 And truck drivers and inventory cycle counters and tax prep specialists... And robots keep improving, although less discussed on this channel, and there go shelf stockers and garbage collectors and assembly line workers. But not this year probably.
Are there no open source training pools?
I've been wanting to get into programming forever and it's papers like this that's starting to push me over the edge. I figured it wasn't costly to use these AI's, just time consuming if you wanted extra training data that your project would need.
@@jackryan529 There's Lambda and AWS etc., but there's just no comparison. These are the smartest researchers in their field, and the compute power they have access to means they can iterate and experiment with training 100 different models in the time it takes me to afford to train 1.
I was happy messing around, and I even used it a bit at work for some label readers, but... nothing that could replace tens of thousands of illustrators like this.
(I'm talking about developing AI, you might mean something different.)
Hopefully, Open-AI will keep their word and make the code opensource for everyone to use.
This is probably one of the most impressive results I've seen on this channel. All the other papers which covers AI generated images are always just that little bit away from being really good. I think this one just blows everything else out of the water in terms of what it can do and its end results - at least as seen in this video.
In a few years we'll have AIs that generate award winning films from a simple one sentence prompt.
And video games
I expected this kind of stuff... eventually, not basically out of nowhere like dall-e
It won't need a prompt.
maybe they already are doing that, this is just the technology that is public, imagine what we don't know about
Or we'll just type "space adventure where ragtag team of misfits visits a different alien colonised planet each episode" into Netflix-dalle-vision and watch what the AI creates specifically for us....
That actually feels possible after watching this, my mind is blown
This is way too advanced for it to even exist...like WTF?!?! it can do shockingly advanced and complex stuff, this is like taken straight out of a scifi movie (like something from the far future) but the crazy thing is that it is real and already exists.
@@1seb11 I already heard about these.
“we overestimate the impact of technology in the short-term and underestimate the effect in the long run.” For this reason, variations on that phrase are often known as Amara's Law.
Great video, but I wonder if you would reference the new "Chinchilla 70B" which greatly outperforms GPT-3 175B, Gopher 280B and other large models?
Sadly, it's already obsolete in terms of bleeding-edge tech. Google's PaLM outperforms Chinchilla.
@@IceMetalPunk But it's much larger. Chinchilla is important because it showed we need to scale the data and models together to extract all of the juice. Chinchillla is parameter for parameter no. 1.
I am struggling to believe my eyes - how has there been this much improvement in just a year? dall-e was great for inspiration but created very dreamlike images, where bits of the pictures sort of look right but you can't tell exactly what they are representing. These have a scary realism to them, like you could just generate them and immediately use them and people couldn't immediately tell they weren't real. I'm honestly glad it is still struggling a bit with the lettering!
If you watch closely a lot of details are wrong, and it’s not always clear which images have been altered multiple times
@@hombacom the images aren't altered. The AI churns out a batch and the researchers chose the best ones.
@@fedyx1544 there is an inpaint tool where you can replace specific parts of the image with new content. Because it’s not available for public people make their judgement of limited amount of images and not the actual experience.
@@hombacom "...it’s not available for public..." *Yet. They're rolling out beta access on a staggered waitlist model.
@@IceMetalPunk How about someone actually demonstrate the service live so we can get an understanding how it is to work with
I'm excited for when we can use these to generate complex textures for 3d models. A tool that can generate from the target geometry and text, and let us adjust while it keeps track of the context would be so helpful! Even something simple like stating "Chipped paint on metal with some rust", could give us slider controls for those context to tweak the output further.
Has anyone ever asked it to do a self portrait? Absolutely fascinating, the progress is amazing. What a time to be alive!
extremely impressive. and this time, genuinely scary. never thought an AI will reach such levels in my lifetime, let alone until i reach my thirties. yet, here we are.
what was that scene... "can a robot write a symphony? can a robot turn a canvas into a beautiful masterpiece?" well, now it can. for better or for worse.
*and
"Can you?"
It cant write a symphony yet. At least not a good one
Worse yet, look at how much data people are sharing about themselves online. I'm scared to think about how future AI could be used to manipulate and deceive.
But most are dedicated to only one task, there is no semblance of general intelligence, that is a far more complex problem(if it is one that we want to solve).
This is insane. I thought we had more time until we figure out AGI. Chances of solving the alignment problem until then seem remote.
just few more years my friend.
@@mr.sunflower3461 200 years?
We can see that this AI can "creatively" improve existing images bounded by parameters supplied by the user. All that needs to happen for AGI is for AI to "creatively" rewrite its own code bounded by parameters that "improve its code". It still sounds very abstract, but so is understanding context in images. We are some years away regarding using AI to improve source code, but once it starts it will kick off very quickly.
If AI stays like this we may just avoid the alignment problem entirely.
@@TomFranklinX Could u elaborate plz? If it stays as a kind of oracle and non-volitional we're fine or what?
Dall-e 2 is the moment I stopped doubting AI, this is INSANE! It's so hard for me to believe that its not just finding stock photos and combining them + applying filters.
It is kinda like combining stock photos. Except it has been trained with millions of them. Plus, to give a convincing photo for a given prompt it has to "understand" language and what words look like, which it is not easy.
Even if it is "combining" photos, it is amazing that it can do it in an almost human level for every possible context
@@albertoaltozano8354 What do you think your brain does. All the things you imagine come from the things you once observed.
@@albertoaltozano8354 Well, kind of, but not really. It's not so much combining photos as it is combining visual concepts that it learned from photos. Small but important distinction. And as May the Science be with You said, that's basically all our brains do, too, when we imagine things.
The question is... if it was trained on other people's material, whose is it? If it was trained on copywrites, should they not own a small piece? And if it was trained on the public and free, shouldn't it too be public and free?
@@WaynesStrangeBrain Should Steven Spielberg own a piece of the copyright to any movie that was inspired by Jurassic Park?
I think this technology will be essential in order to solve full self driving. This model has a deep understanding about textures, objects, patterns, and how they relate. It can imagine things and complete scenes with a lack of information. We do this every day while driving. We don‘t just look at the road we see, but we imagine all the road and objects around us, even hidden ones, like behind a corner.
if i have such software , self driving shit would be my last or one of the last choice to use it on lol.
But the problem is it is insane computation heavy. Unless you want to drive around with a supercomputer attached to your car with an energy consumption of a dozen households.
The software is improving faster than the hardware right now. On the other hand, train GPT-3 to design hardware or train it on every scientific paper / concept whatever. Damn this is fascinating but a bit scary too.
@@prometheus9096 Exactly, it is computation heavy and we have exponential growth in machine learning, it's what makes me optimistic and pessimistic about full self driving. It can or can't happen anytime 😀
Wow, I cannot wait until tech like this makes it into the main stream. The first thing I would use this for is a surreal, unique DnD campaign!
"Armor plated shark jumping out of desert dunes"
"Birds eye view of haunted forest castle"
"Dall-E self portrait as a Wish Dragon"
I can imagine better uses, but this would be too much fun.
You know that story about how when the first motion picture was shown to the public, where a train was coming towards the camera? It's said that people started running out of the theatre in fear, because they truly believed a train was coming towards them and was going to burst out from the screen. I feel like those people right now. I cannot believe my eyes
once again extraordinary. thank you for bringing this to a much wider audience. can't wait to use (something like) this to develop animated characters and look-dev as well as mood boards and (perhaps with pose estimation?) even storyboards and animatics. let us hope that publicly available DALLE/2 alternatives (perhaps based on GPT-NeoX?) arrive soon!
You commented this hours before he published it?
17 hours ago?
this was uploaded a minute ago
@@windy3854 Patreon supporters duh
@@santiblocks I forgot about that, feeling dumb now
Dall-e 2 isn't GPT based so the public models wouldn't be either. People would probably come together to try to use the LAION 5B dataset to train a CLIP model and then a public Dall-e 2 based on that.
I have never wanted the source code and models for an AI more than I want this one, the possible uses are absolutely astronomical.
I don't think it's ever going to be released. This is a trillion dollar technology.
@@motox2416 They cite the reason is "bias" because the AI has a tendency to default to white dudes, like who tf cares, that's something that can be worked on and improved with time.
I'm starting to wonder if Károly himself is in fact the AI voice output of an AI video content creator with the mandate to either make us friendly toward AI as interspecies dialogue... or to lower our guard down for his skynet finally rises up.
The Two-Minute-Papers Basilisk.
It judges your future based on your past actions-if you held on to your papers or not.
agree with you that the voice in this video appears to be computer generated
What a time to be alive! When machines are replacing even the most talented humans overnight. Wow!
@@socialsiege it's only a real proper Hungarian accent
I truly believe that this will even get to the point of making whole photorealistic VFX animations. Artists like myself have our days numbered I have no doubt, eventually anyone will have amazing art at their fingertips rendering human professionals redundant.
I feel like it'll at least speed up the process. AI is a bit iffy sometimes, and isn't as perfect as a hard, cut-and-dry algorithm. I think it will be used for a long time as enhancement to the workflow, to mimic the end result of a fully detailed render and help artists get a prototype visual faster than a full render. It might also help to clean up bad (rushed) VFX work, but there's only so much you can do with that before people criticize you for doing it.
This is the most incredible and groundbreaking thing ever to happen in AI. Beyond words!
There is a very interesting discussion on the project's website about the risks and limitation of Dall-E 3, this technology can be very disruptive and I am happy they are aware of it, but the general public rarely is concerned, and the AI enthusiasts are pushing too fast. This is the time to bring the ethical, economic and creative aspects of AI. I really would like to see a 2 minute paper talking about this.
Definitely something that needs to be brought more attention.
Finally, someone said this, this AI clearly uses data set sourced from the net that also contain works of creators and such, whose work is directly responsible for the output quality of this AI, there's some serious legal discussion to be had here...
There are many papers and textbooks on the ethics of AI if you look for them. I am quite sure that researchers and companies directly involved with this are aware of these factors.
@@diogeneslantern18 Researchers? Sure, companies? Not so much. Whoever pushes it first and is ahead, makes the bank. The thing with AI safety is that all the stuff needs to be carefully and slowly be taken into consideration, getting the first one to publish a paper disregards that basic principle, look up Robert Miles on TH-cam, even that guy seems quite naive and overly optimistic about the whole safety concerns, the future sure looks bleak with this kinda tech in the hands of megacorps.
I agree! It's terrifying to know that the graphic design degree that I just obtained will eventually be completely obsolete.
Finally my dream of being able to convert images in my head to drawings in an instant is becoming a reality.
Scientist said that creativity was going to be last thing AI was going to takeover 🤦♂️ this is really crazy!!
But it isn't creative. It still needs a prompt to come up with this pictures. If it would draw it all by itself without any input, you could call it creativity.
@@prometheus9096 I know it’s not really creativity yet, but eventually it’s going to get there with out needing a prompt it’s just magical times we are living in
@@CrabTheSecond And that will need a prompt or input too. At max you feed it with noise, but thats no creativity its random generation.
Its a fricking tragedy. Humans become more lazy, less determined and less creative and focused by the minute because of all this ish. We will end up like the people in WALL-E.
If the difference between DALL-E 1 and DALL-E 2 is this extreme then I can't even wrap my head around how much better 3 will be
Videos?
@@johntheux9238 or 3D models.
Better eh? You should be terrified at what the next generation of AI will bring. We've been warned by the brightest of minds for decades what will happen if we push the development of AI too far. The complete and utter obsolescence of mankind. Have fun being in awe of the beginning of the end.
@Nick Parsons That's a bit alarmist isn't it? People see too much "Hollywood" and then expect that to be the only future possible. If you cannot fathom how it could help, imagine being free from crushing work day in and day out and instead being able to pursue your passions. That's the future I see, one where AI aids research, we have perfect bodies due to genomic AIs and we can experience things beyond our current imaginations. That's the future I hope I can live to see.
I think this is my winner of the year paper, insane!
And i'm a bit proud that i perfectly guessed the prompt at 9:34 :D
Many of the techniques posted on this channel are impressive, but this one just seems like magic.
Open AI streaming service coming in 5 years where AI generated Shows will be available and you can specify what exactly you want to watch.
Yeah I've been telling my friends that this would be a thing for a while. Eventually we will have shows that are ai generated and possibly catered to individuals... which you have to wonder.. would that even be a good thing?
The holy grail would be ai generated video games. Imagine characters and a world that is rendered in real time by an ai. Imagine infinite open worlds with growing trees. Imagine video games without jagged geometry or texture seams.
Probably the porn version will get created first, that's the way things usually go 😂
@@kennygroth4938 One of the best parts of video game and shows is discussing them with friends and family and lovers. If everyone gets their own it would not be possible.
@@fedyx1544 yeah exactly. You have to wonder how that would go down. What would people overall prefer... I mean if we literally have the ability to say "make me an entire remake of GOT S5-7 that stays true to the earlier seasons", and an ai will make it, it would be hard for anyone to compete with that.
Imagine the power an ai would have in CONVERTING BOOKS TO MOVIES.. that seems like a for sure future
@@kennygroth4938 Imho the media landscape will not change that much, in the sense that AIs will constanlty churn out movies, anime, tv series, vidya etc, then the most dedicated/addicted humans will experience a LOT of them and tell the rest of us where the AIs really surpassed themselves, and those shows will become popular. But eventually we could get such personalized experiences you mentioned
Dall.E 2 seems to be like some alien technology. Wow. This is so freaking mind blowing. I bow down to this research paper. 🙏🏽🙇
The composition and layout of the images is stellar. Wow. I would have expected it could make a perfect image of an item as that is often the negative side of electronic images is that they are too perfect and not realistic. The placement and layout is so pleasing but the more we know about AI the less we should be surprised because if a human can learn all of these skills why not an AI?
I can imagine a subscription service online where you can request fantastic images and have it create them for download or purchase on printed canvas etc. What would be creepy is to request a self portrait of yourself at your time of death. (Scary Music Playing) and see what it comes up with. Once AI's reach singularity they will be able to do this for you.
This is the most incredible piece of software, I think ever. I'm on the wait list and it's going to be better than Christmas
I thought that it was hilarious that I had to complete a captcha to confirm that "I am not a robot" to submit the wait list form.
I need this in my life, I'd decorate my entire life with art based on exactly what i want
you can commission artists
@@juhotuho10 but then he has to pay lmao
@@juhotuho10 yeah, but commissioning a artist, paying them a ton and being like, meh, don't really like it, try again, that's too expensive for a desktop wallpaper
@@MrTurbo_ That is why in a proper commission the artist will send you sketches before rendering the final piece, I am sorry to hear that you commissioned an artist before who didn't even send you sketches and the process.
30 years ago people were impressed with this newfangled thing called the internet because you could send a text message instantly to any place in the world. How amazing! Those people could not foresee that today the internet is an all-encompassing part of our lives and we can't imagine living without it anymore.
We are at that moment in time again.
The funny thing is that for the past three years, I've suspected that AI technology approximates subconscious thinking, or perhaps dreaming. Now, we see an AI struggling with writing text in images, which, to humans, is a hallmark of being in a dream.
I had that feeling when watching the video but you put it to words very well!
great comparison
This is the beginning of AGI's awakening
I’ve been thinking something similar, but you put it better!
When we get a video version, be sure to ask it to flick a light switch or check the clock.
Everytime I watch a 2 minute paper vid I'm exited and scared at the same time , the exponential advancement is crazy ! We are creating a god !
Yeah i have the same feeling, perhaps not a god, but someone better then me (us) for the first time, we enver had that experience on this planet being the most dominant species, never knew what its like to look up to some superior creature that just knows how to do things that are beyond our understanding like how lots of anbimals must experience us, we will get to experience that, but at leatse we'll also get the benefits of their generosity like animals get our benefits, they could take us to Andromeda for example, the ways of which are just beyond our understanding or even hoping to explain to us but they could do it.
The answer to "what would you use this for?" is "anything". Asset generation for art, video games, animation, websites, apps; the possibilities are endless.
Yep. Be lazy and let the machine do everything for you.
@@danshakulawrence983 Absolutely! That's why humanity invented tools 😃
In reality it will end up being used mostly for weird fetishes
This could possibly be - I don't even know. When this hits the mainstream, something is about to go down. I can't wait to see it
@@katech6020 That would be the good outcome, as long as people aren't hurting any (real) people in the process or generating anything outright illegal like content involving children. Unfortunately, OpenAI currently bans any adult, violent or otherwise potentially controversial content.
I genuinely believe at this point that we will live to see sentient AI.
Imagine 10, 20 or even 30 years from now.
lmao slow.
Probably not even in the next 50 years.
@@kombinatsiya6000 is there a reminder thing on youtube? I would love to come back to this in like 5 or 6 years and gesture vaguely at things :P
I agree, this is now within our life time. Look at this machine! Look at what it is capable of. Exponential curve of hardware speed is flattening, but there is enough juice in the tank to get us there imho.
It’s supposedly going to be a thing by 2030ish.
That's it fellow artists. we're done.
The age of AI has begun.
It won't stop with artists, it comes for everyone, including these naive idiots worshipping their replacements
There will still be a need for art managers and art directors. And artistic expertise. But ya- manual painting I think is gone in 5 years
@@dariofromthefuture3075 I think that prediction doesn’t account for the appeal people find in art being produced by other people
@@hungrycrab3297 Agreed.
@@hungrycrab3297 Stock footage on the other hand, doesn’t need that personal touch and I don’t find it hard to imagine its total replacement with AI
This is growing and growing and no one I know will listen to me about it. They are going to be very shocked when they start to see the results in the future, its impressive enough to us who can somewhat grasp it.
it's scares the f#!% out of me! I started a new computer laptop I got starts in seconds... it plays games it shouldn't be able to play it loads actually faster than my real computer...
This would be revolutionary for gaming. Imagine a video game where this AI could generated new outfits, cars, ships, items, etc., autonomously. For example, lets say you are playing Star Trek Fleet Command. You want to build a new ship that is unique and different from everyone else's ship. You could tell the AI to generate a "borg ship shaped like a pyramid with green glowing lights." Or "USS Enterprise with 3 warp nacells and cloaking device."
Or lets say you are playing a MMORPG and want to buy a new outfit. The AI can automatically generate an infinite number of unique outfits for you. Go to the clothing store and say "blue swimsuit with shiny reflective plastic." Or "black thong with leather bra." The AI could generate an infitinite number of custom made outfits for your avatar to wear. And it would take zero developer time since its all AI generated for free. This could allow game developers to add AI generated content like outfits, ships, cars, items, etc. at zero cost.
Instead of spending millions of dollars and years developing expansion packs, this AI could just generate new content automically in seconds for free. This would allow game developers to make new game expansion packs and updates in just days and provide them for free or extremely cheap.
At that point, people would not need commercial games anymore. Just boot up the AI and have it design your entire game for you, and boom, content served. Anyone can have a locally-tailored experience for themselves only. Somehow, this feels incredibly sad and lonely.
Yeah I was thinking for multiplayer shooter games, each arena could be generated before the match, so no player would have advantage by knowing the map beforehand.
I'm speechless, this is insane... the AI has a real understanding of the images, how can a neural network be so expressive? They're just a heap of weights and biases after all, no?
This is looking more and more like black magic
a huge* heap : p
Like 3.5 billion parameters is pretty ludicrous
"They're just a heap of weights and biases after all." Well, we're just heaps of neurons and synapses after all 😃
@@IceMetalPunk ...and working in the same way ....
@@mirek190 More or less, yeah. Obviously there are *some* differences, but overall these networks learn in similar ways to us meat-bodies :)
It uses fractional calculus just like every other artist, obviously.
I imagine this would be helpful for learning foreign languages. A specific text in a target language would have a specific image, which would help with comprehending the meaning of the text (without translating it).
And then we could invent a universal language capable of crisscrossing every past n future languages
I feel like that'd break down even earlier than conventional machine translation since a lot of things you can express in language can't be expressed easily in images. (Everything that isn't just a description of how things are (right now).)
@@cobalius Esperanto is what you're looking for.
@@talkysassis Being such a "universal" language as to be written in the Latin alphabet (not native to billions of people) and composed of vocabulary singular practically exclusively to European languages, it's fair to see why Esperanto isn't quite convincing enough to most everyday people besides those of us in the West. There shouldn't be a need for universal language anyway, because AI is evidently so advanced that it will soon be capable of trivially translating most thoughts, which means we can each use our own languages and retain our own cultures, and forego the effort of learning an entire auxiliary language solely to communicate on an international level.
@@spaghettiking653 That would be amazing
Wow insane, this can kill the stock photo industry, i can't wait for them to release a program for the general public
They're already rolling out staggered beta access. You can sign up for the waitlist on their website.
the public won't get it first. it's a lot more profitable to have it SaaS
@@IceMetalPunk Have you got the access yet? I've just signed up.
@@WilliamParkerer Nope, still waiting. But I'm just a nobody who's spent far more time playing with GPT-3 in the playground than actually using it in any projects, so I expect I'll be in one of the last groups to get access to DALL-E 2.
This is absolutely incredible! I cannot thank you enough for covering this. I honestly don't know where else I would see this sort of news
I was so excited to watch your videos when I first discovered your channel because there were so many things I didn't know about. You inspired me to pursue these fields of interest, and now I'm glad I saw Dall E before you posted this video.... but you have a unique touch to your videos. Thank you very much, Dr.
As an artist this is the most terrifying thing I have ever seen- it was easier to believe that human made art was irreplaceable in the last paper, but I am not so sure now. Scary stuff, might have to reconsider my career choices real soon.
Creativity will become way more crucial than skills by themselves
The amount of time it would take a human to oil paint some of those paintings is probably in the hundreds of hours. While the video didn't mention the time it took to generate these images, this AI can most likely do it in a fraction of the time, and in the future it'll be instantaneous.
@@ryans3979 apparently it's around 10 seconds, but I could be wrong
It's not scary. That's silly.
@@stingingmetal9648 Its not really. It takes people years to be able to draw and paint in the level the AI shows it does. All you need to know is what makes up a good painting and maybe know how to edit those results and viola you are an amazing artist. Way more easy to do.
No way! My brain refuses to believe. This is way too perfect to be real.
Digital artists gonna have a hard time in the future.
As an artist this is gona impact us badly a lot.. if it can do this much i dont think most of us can complete with it..
when the tech gets better id like Dall to interpret my poetry
Man speaks with a full stop after every 3 words.
Idea for next paper: DALL-GIF - Creates short gifs based on user prompt. It needs to now understand motion instead of just a single image.
that would be the dopest thing ever ! I actually think it will happen sooner than later.
Amazing! Imagine this with a 3-D engine coupled with programming AI, VR, haptic feedback -> Holo Deck -> neural inputs -> Total Recall/Matrix?
I really want one of those avocado chairs...
This is mindbogling, art is changed forever
That picture of self portrait touch my heart.
It looks sad and overwelmed.
We ask them to do a lot for us. But do we ever asked what it wan't to do? What are it dreams?
Maybe it's not how it see itself, but how it want us to see it. It's a cry for help, a message.
It's crucial to decript the message. Because it can go from seeking empathy to seeking revenge.
I wonder how many more papers down the line it will be before these kinds of AIs are making feature length movies or bingable tv-series.
“Can a robot write a symphony? Can a robot turn a canvas into a beautiful masterpiece?”
literally both now
Open AI needs to release this to the public.
Open AI doesn't like to be too open
They are. They're currently rolling out staggered beta access. You can join the waitlist on their website.
the most amazing thing to me is that the teddy in the top left result at 4:40 is even doing a freaking no-comply
One day we will say I want a movie about..... with these actors...... GENERATE!
OpenAI is defined as a no-profit organization, but all code for most of the interesting stuff is closed, or by invitation only, and their services have pricing that says it's not free.
To me, it seems they should be called more Semi-OpenAI, suit the name better.
ClosedAI
The fact that these results seem too amazing, and hearing that they are *NOT* sharing the code, makes me think this is actually a fraud.
LMFAOOOO
@@MaxSMoke777 I'm confused... Isn't there a peer-reviewed scientific paper associated with it?
SellOutAI
I’m excited for when this reaches full video level, and we can basically request the AI to make customised films for us.
I'm not excited at all for that. and tbh I have absolutely no idea why a person would be excited for that. "yay more artificial garbage for me to waste my life consuming"
@@enei7045 There are many people who like to spend time in creative pursuits, and this will unlock a great deal of creative power for people of limited means. Other than that, if you don’t like games, films, novels, music, etc. and consider such things a waste, then congratulations for having an opinion no one cares about.
it'd be very interesting to look at the images the AI references, or perhaps do a reverse image search on it's output to see if the output is a manipulation of an existing image.
Yes, I was wondering if it was even able to list the images used for inspiration. They might have simply 'soaked in' to the NN and have no reference back to the originals.
Would the capability be baked in if such a reference could lead to claims of copyright infringement?
I certainly have a lot more to learn about it...but I believe it's essentially like @Octamed said, and there is no direct reference to any existing images inside the "black box" of the neural network. Not that there _are_ references and we just can't see them...rather, the training data imageset serves to fertilize the AI with a vast network of text/image associations, and the generated image is pulled from this web of associations, not from any reference images. But I could be wrong! Especially in particular cases where a text prompt has very few examples in the training imageset, it could be possible there's a source image with qualities that get propagated to the final output image in an immediately recognizable way.
Those parameters aren't storing any actual image or any data at all. It's just like how our brain conjures up images. We see the image and it's being conceptualized into the model.
I have early access and many of the generated images that it makes have remnants of shutterstock and other stock site watermarks
I thought you, as a light transport resercher, would show the pictures of glasses of beverages with sunlight passing through them. They are beautiful.
All of the papers you cover are excellent, but this one is the first one in a while that really blew me away. Out of this world, huge congratulations to the authors!
I am surpassed. Amazing! The AI is better at art than I am. I only hope other people will want the human touch. Top right of the second set of photos best represents me holding onto my papers!
I chose top left on the second. 'ey neighbor!
@@ChuckSploder that one was also great!
I have a feel that dall-e and its variations will each have a certain feel that we are going to be able to recognize. At least to people with more trained senses.
We need to have a conversation about how to utilize this technology but also preserve digital artists. They rightly feel threatened by this. I don't want to lose them!
You can’t. I’m sure there will still be a few people here and there doing things the ‘old fashioned’ way, much in the same way there are still people learning to oil paint now, but they will be something of a novelty and there won’t be studios hiring digital artists to work on their film, game and book development pipelines in any significant numbers. There won’t be people willing to pay a fair amount for digital commissions except for the kind of useful idiots with too much money who pay for NFTs now. I have a feeling that by the time the next century rolls around all forms of art will be regarded solely as something which spontaneously generates without outside help for our passive consumption.
This is the most incredible AI I've ever seen. And I watch a lot of this chanel haha.
The basketball player oil painting one was seriously impressive, its truly expressive.
I can understand artists and graphic designers feeling apprehensive about this.
But as a creative with absolutely zero artistic skills, the idea that I could put my ideas into words and they could be realized through this method makes me very emotional. It feels like I've never been able to express myself and this would let me.
I think this could be a great way for artists to get inspiration for their projects
Lol soon artists will be redundant
no offence but that's so naïve. this will replace artists completely. also that's pathetic to think an artist should get their inspiration from a fucking AI hahahahahaha. fucking hell this comment section makes me weep for the future of humanity. we're just pathetic unaware animals doing as we're told.
@@enei7045 Who hurt you?
@@ferench1145 the people who think AI will improve anything. industrial technological society and its consequences have been a disaster for the human race.
I hope that this is another step on the path towards neural rendering for games
Imagine combining a text-based language model like GPT- or PaLM with a state-of-the-art voice synthesis AI (is Tacotron 2 still the best in the field? I haven't kept up) and a DALL-E successor that can keep temporal coherence to switch viewpoints, plus a NeRF approach to convert those to 3D. And all on future improved hardware so it can be done in real-time.
@@IceMetalPunk if this gets real in the future and reasonably optimised for an average graphic cards/CPUs like for instance DLSS, its going to be awesome! Devs will just have to concentrate on optimising game code instead of graphics
This 100% marks the end of the Photoshop artists' careers. Kind of disturbing, but at the same time amazing how much you can do with the software these days.
This will only empower artists. You don't have to wait for things to render. Your ideas can just come to life. No waiting for your vision to become a reality.
@@theplayerformerlyknownasmo3711 ideas are gonna be worthless when a computer can make sicker ideas. I bet computers will be making arts better and more moving than anything a human could even dream of.
One guy writing a few words can now produce the same output as 100+ artists. This means 99/100 lost their job :/
Still takes an artist to know what's good and what's not. An executive can still request bad art form a photoshop artist, and half the battle with the client is usually trying to convince them that the better idea is their idea, so that they don't get offended and/or stick with the bad design. This also means that... AI isn't going to be able to top us artists yet xD
At best, they can make some cool NFTs... or crop that one guy out of the background of your family photo.
@@oholimoli not even hundreds, far more. Waaay more, it's absurd how much work this bad boy can churn out instantly. This isn't including the skill, or lackthereof, from the person judging if it's a good image or not. I really see this as a tool for artists rather than a tool to get rid of them.
lmao, the oil paintings of peeps holding the papers were a high-level meme xD
What a time to be alive, indeed! :)
I'm imagining video games mostly generated by AI. People, Quests, and places constantly evolving in a world as complex as the real one. What an amazing world we live in!
also, real time AI video generation? I could see it happening
Great video! thanks again for awesome content :)
This my friend, is an interesting future :) Procedurally generated quests with GPT like machines in the background
Based on what is required for images, I'd say that we're probably at least 50 years away from that. AI-assisted creation of games, sure, but AI _driven_ games (sans proof of concept "games" such as AI Dungeon) would still be too demanding unless we get a paradigm shift similar to when we went from vacuum tubes to microprocessors.
I'm an artist. This is both stupendous & bloody frustrating.
Could the same AI be used to analyse images for evidence of photoediting?
I believe there are systems being developed which are created specifically for detecting deepfakes/AI-generated content. Though I'm missing a name right now...
At some point photoeditting will not be able to be seen as edited photos.
@@fiso64 Maybe the real question is how long will anyone even care? Who will be making anything "real" when this, and the coming iterations, are your competition? People have less willpower than they think, especially when things are profit motivated. What about once the images and words are optimized for virality specifically, not just being passable as good art? We saw what those kind of algorithms did with social media.
@@fiso64 Thanks for the good thoughts. Nice analogy. Another thing is we honestly don't know how integrated this technology is into the media we already consume. Why does it take 17 writers to make a hit song? I get really concerned with gross appeal that is just undeniable. It feels like we're being hacked. Marvel is the perfect example.
Maybe the A.I. will create nature preserves for us ancient types. Maybe that's where we already are, but we're ruining it. Maybe we've got one last creative round worth generating and then it'll get reset. That's quite a myth to ponder, but it reveals some things about reality that are certainly true.
I say that, because from a broad point of view, you can't just create art. Something else must be created that has the "soul", if you will, to comprehend it joyfully. We have such trouble appreciating our humanity.
Is there a playlist for these 'legendary works'? I definitely think this would qualify. It would be great to have a best of the best video list for showing friends and family the channel.
Digital jobs will be hit very hard with technology like this. Give it a few years and 70-80% of designers will be replaced by a few people people working together with AI.
Yeah that's a huge concern I been stressing about for a while ... I don't see a good solution this will fuck a lot stuff up ...I mean I hope I'm wrong and this thing doesn't prove as good as we might think in general use but so far it seems preett terfying ....
@@HumanTouchArt Gouvernements will either have to give companies extra funds for creating more jobs for humans or there has to be some kind of unconditional basic income for anyone when more and more people get unemployment to prevent revolutions. One way or the the other that money has to come somewhere and we can only hope that it will be taken from companies over higher taxes. But yeah it’s pretty unrealistic. The future is a powder keg for all kinds of fucked up scenarios.
@Nbsthj Because you don't care about people losing their careers in art? well, I personally, as someone who loves art think this is a very sad moment for artists.
Stock photo industry dead.
Graphic design industry almost dead.
@@digitalsamurai42 yeah basically we it was fun to deam of being a graphic designer or any type of 3d creative job 🥲
The final image of the flamingos had a reflection in the pool.
Its great but it really needs to be more "open".
You ask "What will you use this for" but the first question should be "will I ever be given access".
I think many will share my concern that it does not seem likely with OpenAIs track record so far.
I am a person that is actively engaged in computer graphics as an artist and I love the potential of this, but as I said, so far, I have my reservations for if I will ever be open access.
Previous diffusion models, ones not from OpenAI, but from other scientists, are truly open and one could easily access and run notebooks with. I hope this is the case with this at some point, and I hope its not gate kept behind a big price tag (one unrealistic for an individual to pay, but aimed at corporations).
Bingo. This will be a tool of big industries to cut costs, rather than helping small creators or "democratizing" art. Whenever you hear "democratic" or "democratize" just run. It's almost always euphemism for something bad.
@@LuiZ-jy1pi ”we're bringing democracy to Iraq“