OpenAI’s DALL-E 2: Even More Beautiful Results! 🤯

❤️ Train a neural network and track your experiments with Weights & Biases here: wandb.me/paperintro
📝 The paper "Hierarchical Text-Conditional Image Generation with CLIP Latents" is available here:
openai.com/dall-e-2/
📝 Our Separable Subsurface Scattering paper with Activision-Blizzard:
users.cg.tuwien.ac.at/zsolnai...
📝Our earlier papers with the caustics:
users.cg.tuwien.ac.at/zsolnai...
users.cg.tuwien.ac.at/zsolnai...
Try it out: www.craiyon.com (once again, note that this is an unofficial and reduced version. it also runs through gradio, which is pretty cool, check it out!)
❤️ Watch these videos in early access on our Patreon page or join us here on KZbin:
- / twominutepapers
- / @twominutepapers
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Benji Rabhan, Bryan Learn, B Shang, Christian Ahlin, Eric Martel, Geronimo Moralez, Gordon Child, Ivo Galic, Jace O'Brien, Jack Lukic, Javier Bustamante, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Lorin Atzberger, Lukas Biewald, Matthew Allen Fisher, Michael Albrecht, Michael Tedder, Nevin Spoljaric, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Rajarshi Nigam, Ramsey Elbasheer, Steef, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, Timothy Sum Hon Mun, Torsten Reil, Tybie Fitzhugh, Ueli Gallizzi.
If you wish to appear here or pick up other perks, click here: / twominutepapers
Thumbnail background design: Felícia Zsolnai-Fehér - felicia.hu
00:00 What is DALL-E 2?
01:40 DALL-E 1 vs DALL-E 2
02:15 1 - It can make videos
03:04 2 - Leonardo da Apple
03:19 3 - Robot learns a new language
03:28 4 - New AI-generated drinks!
04:12 5 - Toilet car
04:27 6 - Lightbulbs!
04:56 7 - Murals
05:11 8 - Darth Ant
05:23 9 - Text!
06:14 10 - Pro photography
06:28 Subsurface scattering!
07:10 Try it out yourself!
07:45 Changing the world
Tweet sources:
Fantasy novel: / 1527312285152452608
Leonardo: / 1
Encyclopedia: giacaglia/status/...
Modernize: / 1513588042145021953
Robot learning: / 1514217698447663109
Drinks: / 1514735924826963981
Toilet car: / 2
Lightbulb: / 1513837678172778498
Murals: / 1
Darth Ant: / 1519224830989684738
Sign: / 1
Hand: / 1521755209667555328
Károly Zsolnai-Fehér's links:
Instagram: / twominutepapers
Twitter: / twominutepapers
Web: cg.tuwien.ac.at/~zsolnai/
#openai #dalle

[@Hookens]

3:38 I think we know why you're super happy. These light reflections are amazing, and I can only imagine how much more interesting it must be to someone with a PhD in computer graphics.

[@radekmojzis9829]

When i saw those, my jaw dropped and stayed that way for the next 5 minutes...

[@fontende]

@@radekmojzis9829 saw what? First time hearing about marketing and advertising lies? If you can't recreate chemical reaction the research paper of unproven theory can go into bin 🚮

[@thNumber]

i love light simulations too

[@heckYEAHman.]

@@fontende LMAO what

[@radekmojzis9829]

@@fontende my point is - if the images with light caustics were generated by the AI - even if cherry picked i am impressed. And since this is no simulation AI, but art AI it does a really good job

[@libberator5891]

6:40 only a light transport researcher would point out the subsurface scattering and not the extra finger amount

[@MikkoRantalainen]

One could argue that fixing the finger count is actually much easier problem to solve.

[@AlanW]

It looks like there's also a second thumb partially occluded 😆

[@itz_Metaa]

Lol I didn't even notice until I read your comment

[@mrgreenskypiano]

Dangit, you ruined it!

[@AZTECMAN]

@@MikkoRantalainen How would you go about repairing the ability of CLIP to count? It is well known that CLIP doesn't count very well, and DALLE 2 is basically just CLIP guided diffusion. I think you could argue that it is a easier problem, but I suspect it would be an argument based on ignorance. I would love to hear a well formed rebuttal though.

[@BingusDad]

As an artist, this is truly scary. For the first time ever, I feel threatened by AI.

[@spider853]

Just use it as a tool for inspiration for your work. 👍

[@albertoftasmania]

@@spider853 That doesn't solve obsolescence. If it's his job (graphic designer, commission artist, illustrator etc), he's fucked.

[@BingusDad]

@@albertoftasmania If it can do what the video claims, yes, we illustrators and graphic artists are pretty fucked. Oh well. We had a good run I guess ;)

[@albertoftasmania]

@@BingusDad There will always be a demand for humans, but unfortunately I think (take what I say with a grain of salt) it'll be like the horse and the automobile.

[@BingusDad]

@@albertoftasmania We can only hope that you are right. But even "the human touch" is just about finding the right algorithm like everything else. Maybe doing actual physical originals, is something we humans can have for ourself for a little while ;)

[@dinoblaster736]

Combining a future version of Dall-e or Imagen with something like AI dungeon would be incredibly cool!

[@EMLtheViewer]

I believe that Latitude, the company behind AI Dungeon, is already using an image-generation feature for their paid tier of service, but it’s most likely nowhere near the quality of DALL-E 2 or Imagen. It would definitely be cool though.

[@SherifBender]

Yeah but let's just focus on a decent and consistent storytelling first

[@Eschelaun]

@@SherifBender the problem with the storytelling consistency lies in the fact that GPT-3 on the dragon tier as well as the AI systems on the lower tiers use a system of tokenization that are hard limited to 2048 tokens - meaning without thorough use of world info etc every several paragraphs begin pushing information out of the token range.

[@jamesdegriz]

Ive been thinking along similar lines but with the AI generating VR Content via unreal engine.

[@EMLtheViewer]

@@newpro93 The next logical step is just straight-up AI-generated video games with arbitrary inputs. You tell the AI “Make a narrative action-adventure game set in space featuring strategic combat and physics-based puzzles with a vaguely medieval aesthetic, cell-shaded graphics, and a purely a cappella soundtrack.” and it creates an original game to the best of its ability that fits that description, possibly generating elements of it in real-time as you play.

[@DamianReloaded]

Using dalle to generate images from a novel I'm reading is something I'd definitely pay for. Even people who don't like reading could watch and hear novels on an audio book Netflix. It'd be a total success. Someday AI could generate movies from our favorite novels on the fly. This really has the potential to revitalize literature.

[@lingred975]

I'm waiting for an AI to be able to generate a game on the fly for me to play. :)

[@LightVelox]

@@lingred975 or generate new episodes for a cartoon/series so you can watch it forever

[@SherifBender]

To me visualising something that you reading about is like a 90% of the fun that you getting in the process of reading

[@s00ke]

@@SherifBender having AI generate the visuals for you is lazy. Have we become so boring that we can’t use our own imagination to create magnificent picture in our heads? The fact that AI can create art, however soulless, is scary.

[@hyy3657]

All by AI? Nah, more like an assistant of the artist.

[@lfdhrgasdf7579]

With everything I’ve seen with AI this has to be the most important and breathtaking

[@sbalogh53]

Combine this with another AI that scours the web looking for interesting combinations of descriptions and create a portfolio.

[@cenakaze]

6:42 apart from the 6th finger it has done very well!

[@TracksWithDax]

Same thing I noticed. Otherwise, photoreal!!

[@fedyx1544]

no one specified human hands

[@Nnm26]

7th too lol

[@MikkoRantalainen]

@@fedyx1544 That's good point. If you had added "with 5 fingers in one hand", would DALL-E 2 have been able to produce "correct" output?

[@cryspin7]

Polydactyly isn't crazy rare. Feed it enough images and it will see someone with 6 fingers. Is it intentional for the sake of variety? Was the prompt not specific enough or just a bug? I don't know. If you typed "human" what should it give you? A generic white person? An indigenous black person? What about an eskimo? What about an amputee eskimo? That's a human. Should they always be healthy and happy? Hell, I just typed "human" into google's image search and all it gave me was a bunch of fucking monstrosities.

[@RiamuGMD]

aqui enamorandome de las IA, gracias dalasito uwu.

[@Caspian.]

So what if AI generated images become so common online that the AI starts using other AI generated images as reference

[@nickr7090]

A psychedelic experience. 😉

[@Jetsetts]

Pretty sure that might be the first AI language

[@asdfdfggfd]

@@Jetsetts The AI already has symbols it is using. If you ask it to make technical drawings it has developed symbols for length and square that it uses over and over again.

[@MrShaheer]

It already does that. There is a video of an AI that generates game characters from the images it has already generated. I think its on this channel

[@aretwodeetoo1181]

@Ops Blac Not necessarily, because only the good images get published, liked to and indexed (by humans and human-guided algorithms). If we're still doing the curation, AI can use it to improve.

[@V3RTIGO222]

As a creative person, I'd really love to bridge the skill gap in creating art by being able to describe ideas in great detail to an AI until your exact image comes to life.

[@gc8972b]

also a kick in the pants of those that built skill to bridge that gap. suddenly, useless

[@V3RTIGO222]

@@gc8972b Well, artisans still paint real paintings and people stil sew their own clothes even though that was long ago replaced by industry, and people are able to produce art much faster digitally... so did the artists making digital art invalidate the art made by those who paint by hand? The advancement of technology makes accessing multiple fields in a lifetime a possibility and I don't think we should look down upon it because "it took my job", really it just made their job a lot easier. As is said, it democratizes creation, and that in itself is powerful for future generations.

[@RSAgility]

@@gc8972b Artists will always exist. Just because you tell an AI to give you an image, doesn't mean you as an artist cannot improve on it. Thus saving time being able to create many things and choosing the best one for whatever you're doing... Just like machines didn't replace people, simply made the job require less people. Which is good, overpopulation is a major problem.

[@gc8972b]

@@V3RTIGO222 There are substantial differences to these past processes because the artist is not necessary anymore. When Dall-E delivers complex high resolution images made from an individual creative input, and that image is placed on a t-shirt/poster printed by an online service, the creative process is complete. I like democratization, don't get me wrong, it's inevitable and it will improve communication for our species, but the reality, for now, is quite harsh for artists because they/we have been a crucial part of the process of translating creativity for centuries and are now becoming disposable.

[@AZTECMAN]

@@gc8972b Did the photograph render the canvas useless?

[@rogerc7960]

It passes the Turing test for images

[@krunkle5136]

It does at a glance. It needs to get better at close-up details, like mechanical things or eyes not looking wonky.

[@IceMetalPunk]

@@krunkle5136 or hands having the correct number of fingers...

[@juliandarley]

having spent a lot of money on book illustrations a number of years ago, this would have been wonderful. not only that, the price was so high and the process so difficult that it put me off writing more children's books - until now that is. yes, it is true that there may in future be less need for some artists, on the other hand, there may be more things that need illustrating because more people will be inspired to try to write and make films and many of those will still need the special talent of artists.

[@vaibhavkumar-us8tw]

Don't worry you don't even need to write the book, as all the things that you mentioned will be handled by AI at some point in future...

[@fedyx1544]

​@@steorswe they can already somewhat

[@Nnm26]

@@steorswe we’ll see about that :)

[@acaustik8763]

@@steorswe They can already do this, look up free training modules for Novel AI (runs on GPT-3)

[@drorange2261]

@@steorswe Yes I think that with minor interventions you can certainly have Novel AI doing thousand of stories per day for 0 to 6 year olds. But if you train it especially for that task it will become amazing. What is the point though ? We are very far away from having true AI ... so consistent and intriguing art , design or storytelling won't be that easy to replace.

[@dafyddroff8084]

I was so so excited about this when the last video on DALL-E 2 came out. Now I’m actually really terrified of this

[@virtualmartini]

why

[@_festival]

If you're scared go to church.

[@TechnoMinarchistBall]

This is actually pretty amazing stuff. To think that less than a few months ago this tech couldn't even push out something that even resembled the prompts I asked of it.

[@michaelleue7594]

Try a prompt describing a certain number of something, like "a drawing of three flowers" or "a family of four". DALL-E still struggles with basic elements of prompts.

[@CosmiaNebula]

coming next: AI-generated articles arguing that this is not really thinking, just curve fitting

[@MikkoRantalainen]

Make that AI-generated photorealistic documentary uploaded to KZbin...

[@z-beeblebrox]

@@MikkoRantalainen Half the youtube "documentaries" I encounter feel like they're AI generated anyway so

[@juhotuho10]

It is just curve fitting though, AI can't be conscious to argue that it's thinking and conscious means that when i'm solving math on a piece of paper, I'm simulating another conscious being in my brain, and that doesn't make any sense.

[@MikkoRantalainen]

@@juhotuho10 Consciousness is interesting concept and different people define it very differently. A good question to ask is how could you know if a chimpanzee or a mute human is conscious? I'm nowadays thinking that consciousness is just a method of feedback where you evaluate your past actions and to do that you have to understand concept of myself vs others. And I think you need feedback mechanism like this to use unsupervised learning to automatically figure out complex tasks. As such, for my definition of consciousness, future AGI will definitely be conscious.

[@hyeve2114]

​@@juhotuho10 Really depends what you consider "consciousness" to be. If you look at it purely materialistically - without including any kind of religion or faith or idea of "soul" or "essence" in the picture - then, as far as science has been able to determine so far, there's no inherent difference between human consciousness and any other kind of processing. It's just a matter of deciding where to draw the line. When you're solving math, no, you're not "simulating" another being (though, I think it could be arguable), you *are* the being - and there's no inherent difference between the processing that's happening in your brain, and the processing happening in a digital neural network. Neural net AI's, are, after all, explicitly designed to mimic how the brain works. Something that's not super related, but still very interesting I think, is that It's well researched already that our sense of "self" is, in reality, just a conglomerate of huge numbers of different parts, which, when operating normally, all function together. When those parts fail to function together properly, as happens in some kinds of personality disorders, people can and do absolutely end up split into "multiple" conscious beings within the same body, with fully separate personalities and memories and behaviors. It's really really interesting (and also pretty scary, too), and I highly recommend looking into it yourself. Dissociative Identity Disorder is one of those conditions.

[@btarg1]

Dall-E 2 would make an amazing Photoshop plugin for making quick edits or generating elements of an image, or maybe even a video editor!

[@krinodagamer6313]

mannnnnnn

[@USBEN.]

Wait some years to get a system capable of running these models natively and ability to transfer all these databases of AI to personal systems.

[@ollllj]

Designers seriously use dall-E for inspiration, as it has a huge database for style-transfers, so you can easily tell it to render "desktop-sized commodity [A] in the style of [b]", like [table lamp] [seat] [curtain] [neon sign] [candle] [kettle]

[@mickelodiansurname9578]

Problem is the folks that hire Designers will have the same idea... minus the need for a designer obviously!

[@s00ke]

Wouldn’t that eventually make designers obsolete? Won’t they just say “why would I get a designer? I’ll just tell the computer what I want and pick the one I like most.”

[@MikkoRantalainen]

@@s00ke For the near future, the human designers and human illustrators will be needed to complete the illustration if you want really high quality. However, I wouldn't recommend targeting designer or illustrator as a career if you're pondering it now. By the time you're good at it, the AI system has probably already figured out the missing high quality part.

[@hombacom]

@@MikkoRantalainen I thought understand the current product and human needs was the ”high quality part” for a designer and not lack of render quality or brainstorming ideas.

[@ollllj]

@@s00ke dall-e may violate some rules of physics and the designs may be implausible or inefficient.

[@GamingRevenant]

Whilst incredible, this euphoric feeling will only last the first few times this will hit the large public. We will very soon come to a point where so much in every single field is generated by A.I. that we will literally be impressed by absolutely nothing whatsoever anymore. I can understand the potential and the hype, but it simply won't last. I find that to be a very dark perspective going into the future, and it will kill every last bit of motivation to achieve anything for all future generations to come.

[@brandonthesteele]

I see what you're saying with tools like DALL-E quickly becoming dull, but as long as there are humans, there will always be new boundaries to push and new areas to explore, even if we can't see it now. I don't find the idea of human motivation being killed by technology to be particularly forward thinking or imaginative.

[@drawmaster77]

I dont necessarily agree that it will kill motivation. We will just find more complex things to get motivated about.

[@oren2234]

the AI's will crush us at every task no matter how much you train. if you dont find that threatening im wondering what kept you alive until now

[@drawmaster77]

@@oren2234 why should this be threatening? I don't get the sentiment that "hey I want to do this repetitive task myself".

[@awildtomappeared5925]

@@drawmaster77 More complex than art? I look at paintings by the old masters, and I feel a desire to put in the work to try and become as great as them, that is a high af goal that I will likely never reach and it's not simple to get there, but with this tech, anyone will be able to generate better images than I could ever create in just a few papers down the line, this could go for almost any field, this absaloutly kills motivation.

[@aBlackMage]

I stumbled on the subreddit for DALL-E 2 a few days ago and the results people are posting floored me. I can't wait for it to become commercially available so I can try it out myself.

[@tpachalski]

This is insane. The details and the subtle understanding of the text blew my mind

[@BossKillRatio]

I want to see 3D content generation.. would help speed up game development specially for indie or solo developers... and specially those with little artistic ability.

[@alteredcarbon3853]

I was working on a game in my free time but 3D modeling was taking so much time I had to put the dream on the side until we get an AI like this for 3D.

[@johnclark926]

I’d like to see this too, but I’m afraid we’re gonna have to wait a while. Although I’m sure there are many 3D models out there to train on, I don’t think there are nearly as many available 3D models than there are raster images and videos. This means that unless we figure out how to get AIs to learn more with less data, make a crap load of 3D models using photogrammetry, or we somehow exploit the multimodality of an AI to positively transfer knowledge from other data into 3D model creation, the results of AI generated 3D content will be lackluster and not fit for use.

[@blakksheep736]

Can't we combine Ai tho? There are already Ai that can construct 3D objects from 2D pictures.

[@jenkem4464]

@@johnclark926 I think we're closer than you think. Like a previous poster mentioned, there is already AI in the works that can roughly create 3D from an image...and also create really good 3D from just a few images..complete with material reflective propterties. And keep in mind that's only what we've been shown in public.

[@Rybz]

@@jenkem4464 I give it 1 year till we have a text to 3D model AI. I think most of the tech already exists but are distributed by various AIs. They just need to be combined! And I say 1 year because at the speed AI is advancing and with the popularity of DALL-E and the starting of the metaverse, turning it 3D seems like the next natural path.

[@Dheeidjdndbd]

This is so exciting it’s insane. I really want access to this. I used the original Dall-E so much and I got really good at making prompts. I would love to get going on this one

[@alexgeorgiev532]

Would you be so kind and send me a link to test it, please?

[@MattVidPro]

How did you get access to original dall e?

[@Joel-5747]

So cool! Just started watching this year and the presentations only become more and more fascinating!

[@EMLtheViewer]

I think the next natural step in the evolution of this technology is to adapt it for the generation and editing of whole videos. For instance, you could tell the program to add a certain type of hat on the head of a character in one scene (or manually paint it in one frame), and the program will automatically edit the rest of the frames in the scene to have that hat with a consistent appearance and quality. If the program demonstrates an understanding of three-dimensional spaces, physics, and light behavior, then it could become an extremely powerful tool for video editors. For example, it could be utilized to add high-quality visual effects into scenes with significantly less effort than with traditional methods. It could potentially be used to generate entire movies following a storyboard should it prove capable of temporal coherence. Safe to say, I’m very excited for what’s coming in the near future.

[@kakyoindonut3213]

ok that's actually a bright idea, not only can it be a convenient, but it can also be an entertainment itself, imagine how much laugh youtuber or streamer will get from commanding the ai to do something with the video and trying to break the ai.

[@LetsGoGetThem]

It would be cool if a key animator could draw a few frames then have this auto-inbetween it in a rough way, this could make traditional animation viable again.

[@3-design]

Eb-synth is on it's way to do that.

[@hombacom]

It has no understanding, it just replicates what is common in images. Some things work good other not. Have obviously trouble to create things like hands. So the idea that it could do something natural with humans acting sounds very far, because we can easily spot when its unnatural. Two different scenes angles that should be same could suddenly look different, same problem with coloring black and white movies.

[@kakyoindonut3213]

@@hombacom well there's probably other way to make ai understand shits in the future

[@MannyDer]

those murals were awesome! That was such a fast segment too, I wanna see more. Please :)

[@EMLtheViewer]

It would be neat to see a video that compares Google’s new Imagen to DALL-E 2. With two competing systems, it would be interesting to see the particular strengths and weaknesses of each put against each other.

[@IceMetalPunk]

Unfortunately, Imagen explicitly is not planned to be released, so that comparison isn't possible unless Google changes their minds.

[@MattVidPro]

Imagen has basically been deemed too powerful by google for the public

[@ky1ewithsty1e]

@@MattVidPro Corporate assholes.

[@PorousG]

@@IceMetalPunk so what does Google plan to do with Imagen?

[@IceMetalPunk]

@@PorousG No idea. Probably nothing more than just show it off to be ahead in the AI race.

[@AngelooPlay]

Aquí comienzo Dalas, VOY CON TODO AMIGO

[@digitalhouse6969]

the concept of your channel and it's content is rare and really good and rich with info! never stop !

[@imjody]

3:35 - I have to be honest! I wasn't asking myself "Károly, why are you super happy," because I too, was very excited to see the incredibly lighting and reflections seen in this scene, and knew you would be too! 😊 This is all so remarkable!!

[@ArtamisBot]

Imagine a video version of dall-e where it could generate a short video clip of whatever is prompted. You could use that for almost endless b-roll.

[@alucard0712]

I literally still cannot believe in those images. they are so good even from artistic standpoint.

[@Cyndadogo]

Man I really love your videos!

[@lingred975]

this is perfect! I am learning "some" languages and I wrote a few scripts to translate sentences from one language into the others and help me with boring/repetitive stuff. Just last week I was thinking "I wish there was an AI to create images automatically from my sample sentences to incorporate them into my anki notes..." XDDDD

[@thelemoneater]

Low-key considering using something like this to generate prop designs and ideas, 3d modelling them and seeing if I can create a unique environment out of the sets. I don't have an eye for composition, but something like this could substitute those skills while I learn.

[@MrLESLO]

@Two Minute Papers - Thank you for sharing all this, fantastic video.

[@Taylor-made-tatts]

It’s so exciting, I had a dream about this kind of program years ago. I love it. As a tattoo artist this is going to help my reference game like nothing else

[@CinemaoftheMind]

Creating in 3-D programs like unity and blender can do absolutely amazing things if you know all the arcane ins and outs and hidden menus and options for those programs. AI could create a basic mesh object or animation within an entire scene that would still require a special vocabulary for the creator to use it expertly, but there would be so much time saved on workflow trial and error. Exploring a variation or tweaking a texture would just be the right word away!

[@radcliffe2192]

I love that AI knows that human skin have subsurface scatterintg, but not sure how many fingers there are on a hand.

[@IceMetalPunk]

I'd guess it has seen more examples of skin in general than it has examples of specifically hands.

[@MikkoRantalainen]

I would guess the AI network they use is good with textures but not that good with counting. I would assume fixing the counting problem is actually the easy part. If I've understood correctly, DALL-E 2 intentionally doesn't render humans in the graphics so wrong finger count may be intentional, too.

[@IceMetalPunk]

@@MikkoRantalainen It's not that it doesn't render humans, it's that it wasn't trained on photos of human *faces*. They did that to try and mitigate people using DALL-E 2 for deepfakes. So it doesn't do photorealistic faces very well. Hands should be fine; this is just a failure case.

[@xamaximus]

I LOVE DALL-E, i hope we can all play around with a powerful version soon! Thanks for the great video, as always.

[@CamKnoppMusic]

Theres many amazing things on this channel, but this is probably the most amazing one I’ve seem. Usually its like “its cool but I can tell its still a work-in-progress”, but this one looks flawless.

[@AdamSamson1991]

The novel idea is great, and might be worth using for my author roommate. Personally, I would use it to provide visual cues during my tabletop RPG sessions.

[@empanadasdecarne.636]

vengo desde dalas para no morir

[@runningwithSaul]

Nice that the public finally gets a chance for a demo!

[@kylebowles9820]

LOL saving the "what a time to be alive!" for the Toilet Car! He knows a good idea when he sees one!

[@Xavier-es4gi]

This is both exciting and terrifying it really feels like a kind of intelligence

[@onedeadsaint]

this is absolutely incredible! how do we get our hands on the full DALLE2? this will change art forever, and I don't think that's hyperbolic to say that.

[@MartyrMoon]

Never been so early to a 2MP video! This stuff is just MIND. BLOWING.

[@btoe1999]

Just think 5 years from now. Its Dall-E 4 and it create videos out of text and you get an exact replica of the first harry potter book in movie form directly from the text. This is really going to change so much and make our world all the more cooler and more interactive. I love it.

[@cbob213]

We going to talk about that hand with 6 fingers?? I think you were focusing on the subsurface scattering a little too much to see that abomination haha Great video mate. This tech is super interesting to me

[@danmimis4576]

Imagine upgrading from creating images for a novel to creating a movie from a script. Finally my devastating masterpiece, "4 Letters and a Tale" would see the light of day

[@willjames7119]

Thinking more like Spiderman 13

[@DEtchells]

I agree with you, that we will come to think of AI in terms of “before Dall-E 2” and after. That was my immediate reaction when I first saw results from it. It’s not just an evolution of what came before, it’s so much more that anything before it that it’s in a whole new class by itself.

[@peiheng1188]

This is huge and crazy! Kinda redefine what creativity is

[@wellington99930]

Vengo del video de DalasReview, super increible esta tecnología me encantó, no puedo esperar para empezar!

[@teamvigod]

Not one but TWO “What a time to be alive’s!!!!!” Love it ❤️

[@techwizpc4484]

Artists: Fuck!

[@acgeorgern]

I been in waitlist since April! I can’t wait to play with this!

[@briansanderz9024]

This is by far the coolest thing ive seen this month.

[@z-beeblebrox]

As cool as this is, it's frustrating that even multiple months after their first limited distribution to a tiny handful of selected AI researchers, there is zero plan to expand that distribution or even communicate how access to this technology will be granted going forward. It'd be one thing if this was made by a private company, but OpenAI was originally founded as an open source project and it seems they've abandoned that philosophy entirely now.

[@BenoHourglass]

I doubt this will ever see the light of day. There are a lot of issues in the pictures presented if you look close enough. I'm guessing that we'll have to wait for 3 or 4 before they feel comfortable enough to release this to the public.

[@trxps2829]

Yeah, and they didn't even bother to release the original DALL-E, which at this point is outdated. There'll be a new version out before half of the people signed up will even be able to get access to it through their expensive API.

[@trxps2829]

@Poorlord I know, I'm on the waitlist. There are over 200,000 people on the waitlist from what I remember, or possibly more. So most people won't even get access to it before there's a DALL-E 3.

[@trxps2829]

@Poorlord It's not.... Give me a link or tell me where I can find it if it's available to the public like you say. You're just speaking out of pure ignorance

[@DjTechFisch]

Its for a small userbase atm since its still not cleared about copyright (who has the rights to the imagies in the database and copying an artists style like salvador dali, without their consent or without their knowledge since your using their pictures in yiur database is also against copyright) and the fear of people abusing it as fake news, even tho its the internet (they used like a image of putin pressing a juke button)

[@gamerman7276]

Dr. Zsolnai-Feher: "The hands even have subsurface scattering" Me: "So we're just gonna ignore the six fingers"

[@jerobyarts5654]

Holy shit , the images are so good that I feel that DALL E 2 is literally bringing actual photographs from alternative realities.

[@sums001]

That sub scattering surface is so perfect 🤯 .

[@xavieradriaens4411]

Only ia can realy think ' out of the box' i can only imagine what marvelous techniques and possibilities they will be able to show us. Imagine once they will be able to generate solutions with physical theories !

[@chapnico]

I'm watching content about "DALL-E-2" since days now and it is very stunning, I'm french and i'm very frustrated that there's no french videos about it, it feels like no one heard about this crazy tool over here. Really great video by the way, thanks or your great work as everytime !

[@SocksWithSandals]

I've just been trying out the image generator. So much fun!

[@rinner2801]

The accuracy of the lighting is truly astonishing.

[@bustedshark5559]

Thanks for the link to the DALL-E mini. It gives those of us on the eternal waiting list some idea of the potential here. Also, it's very cool in its own right.

[@ziiiboom]

Devoted my entire youth to art, aspiring to be an illustrator for the rest of my life, spent hours and days and months and years studying and grinding through hardship, just to have my entire future ripped by this one AI. When people get their hands on it, especially non-artists, commissions aren't needed anymore. What a time to be alive, now I lost all my appetite.

[@aretwodeetoo1181]

You used to think art is execution, now you've learned it's taste and sensibility. Focus on that and you'll have a job (AI cannot do that and with current techniques it never will). Do your drawings for fun in your spare time. Or do fine art if you're good enough, AI also can't do that. No point crying, adapt (AI can't do that either).

[@alcano6022]

@@aretwodeetoo1181 This is year 3 of Ai generated art.... If Ai doesnt get regulated, doing art as a job will be dead

[@SelfIndulgentGamer]

Those murals are breathtaking :O

[@xavieradriaens4411]

Dalle to make storyboards or inspire characters in game industry would be a must have !

[@Pixelsplasher]

Illustrator / Artist: "There goes my job." Stock Image Company: "There goes our business."

[@Kantirant]

I would use this to do something i've always dreamt of. Create my original comic story with beautiful pictures :D

[@juhotuho10]

you can learn to draw

[@Kantirant]

Unfortunately thats quite hard for me , dont have much mobility in my hands after i had a motorcycle accident.

[@juhotuho10]

@@Kantirant oh, sorry to hear that :/

[@Kantirant]

no need to be sorry !

[@stevenbodai5917]

This is a new brush for artist. An amazing brush!!

[@notgonnabe165]

this makes me happy that things like this exist now

[@Droid_Builders]

I can see a large corporation generating trillions of images and a small artist running into legal issues because they infringed on the corporations catalog.

[@LarsRichterMedia]

And I've still not found anything about how the AI was trained in the first place. What images have been used? Only heard someone say they scanned through Instagram among other places for image-text couples. If that's true, the small artist's work may have been used to build this tool that's then being sold to cororations. Ethically it always seems to be the same old story with new technological developments. Move fast, break things (and people) and don't bother what ethical implications your actions have.

[@dinupetrecristian150]

@@LarsRichterMedia we don't care. The small '' artist'' should quit making art if he is not capable of making a living out of it and get a REAL USEFUL job

[@LarsRichterMedia]

​@@dinupetrecristian150 Who is "we"? The fact that you don't care about ethics and don't seem to understand the issue at hand at all is something that utterly surprises me, random stranger on the internet.

[@crepooscul]

@@dinupetrecristian150 if the job is not useful, why has it been automated and put people out of job. Use your head. Sau mai pe românește foloseste-ti creierul ala ramolit

[@supe4701]

@@LarsRichterMedia He of course referring to our benevolent corporate overlords.

[@free4fire]

Man as somebody who has literally spend years of my life, learning composition, color theory, shape language, anatomy, light behavior on different materials and so on, so that I can paint the images I want, I really feel like an idiot now for wasting all that time (especially since I always wanted to do art as a job, guess I can forget about that now too). I even knew that this was going to come along eventually but I thought it would be 2030 at the earliest but the last 4 years really snowballed with all the A.I. stuff.

[@techwizpc4484]

Don't be ridiculous. If anything this can make man-made artworks more valuable, or at least worth their price (finally). AI-generated work is cheap. This is no different than those NFT monkeys.

[@free4fire]

@@techwizpc4484 Well you're way more optimistic than me, I can very well imagine that in a few years there's going to be an A.I. specifically trained for, let's say concept art as an example, so art directors just need to type in the description they usually give the artist and almost instantly get 50 results back. And besides except for the very few really big name artists companies don't care who paints their stuff as long as it gets done, so if they can get the same quality in higher quantity for cheaper AND faster than of course they're going to go with that. Of course there's going to be a small "pure/true" artists market left but that one will be even more crowded than it is today and even less people will be able to make a living off it. I guess it's time to learn to draw some Furry degeneracy and hope that the A.I. gets told to ignore that side of the internet so that at least that revenue stream remains for a bit. Also sorry for poor grammar, it's 4AM and my brain is mush, so hopefully that mess up there made at least some sort of sense.

[@techwizpc4484]

@@free4fire I don't consider myself that optimistic but at the back of my mind I'm also thinking about those things you said. I mean it's already hard to compete with fellow artists and now we have to contend with AI that can work faster. But if people are willing pay "millions" of worth of dollars in NFT for a cheap monkey, what more for real authentic artwork made by human hands, right? But again people will opt for cheaper alternatives if they can. All creative endeavors will be affected by this but they lack soul, dammit! Us creative people are in deep shit again. But the proponents of this tech say well what about the camera? The reason why artists were so good back in the past was they had too. But when cameras appeared, drawing realistic portraits seem like overkill, unless the client really wants it. I do hope people will still yearn for authenticity. And if I were a popular artist, I'd opt to have my paintings removed from the training data used by the AI. It's just sickening that you can generate an "artwork" in the style of Da Vinci. I think artists should demand to have their works removed from the training data. The only real uses I can see for AI is in the scientific fields but for creative works, meh. Maybe some more open minded artists there will see this as a utility to get inspiration from, It's no different than using reference photos. But with this they can use it to generate something specific.

[@joez6235]

@@free4fire Have you considered becoming an art director? If you think the ai is a better artist than you are, then use it, it's a tool. If you can't get the exact results you want with the ai, you can easilly put your art skills to use and perfect the end result yourself. The reason why companies hire artists in the first place is because they're run by the type of people who wouldn't know what to type into the program to get they image the want anyway. Be the guy, make the art happen.

[@VIN7ERwav]

never had the chance to tell u cause im always watching on my tv but i just love ur videos

[@ClaudiusJulli]

Nice statement: a time before and after Dall-e2. Keep us informed with examples, also about the right text prompts.

[@2k7u]

I really appreciate this project, and out of all others, i would really like to have access to this however most of projects like this get put behind a "waitlist" where most people won't be accepted without reason much like github copilot, gpt-3 to name a few.

[@willjames7119]

Indeed, the elite gap widens. But it's for our safety.

[@Lostintrepid]

Excited to see what can be done with this in the data visualisation space - feed it some data, get an infographic, dashboard or data illustration.

[@elaineyoung1931]

Oh my gosh it looks amazing, my photos that I would normally delete can be changed so it is incredibly usable awesome.

[@xxxvertigoxxx]

Dalle blows my mind every single time I see it.

[@anatolydyatlov963]

"Open" AI who censor their work, don't release the code and employ waiting lists. They should be called "ClosedAI".

[@USBEN.]

Lol true

[@thelonespeaker]

I just wish I could get access to the beta. Been on the waitlist for three months now 😔

[@ZZ-vl5nd]

Me too 😑

[@RussDnB]

This will be an amazing tool for drafting ideas together

[@pietheijn-vo1gt]

This will be an amazing tool to generate training data. I can't wait to get my hands on it

[@isatche]

I would love to see it re-imagining The Lord of the Rings, dune worms or some other classic fantasy/sci-fi literature

[@JoseMorales-jv6fz]

Aqui por dalas, let's go

[@MrFEARFLASH]

I think this is ideal for notes to the script of any video. I quickly wrote what should happen in the scene, and now you have a picture for a team of like-minded people!)

[@Krischenn]

Loved the original Dall E concept and practice- Dall E 2 is on a whole other level 🔥

[@aalluubbaa]

I think the future of television is personalized films or tv series. You just input some parameter like length, genre, style, rating or whatever you want and it would give you a personal tv series just for you. The fact that we have DALL-E2 in 2022, I think it's safe to say that when my 1 year daughter grows up, she's going to watch her personal tv shows when she's teen.

[@hc3d]

Perhaps it could be automatically tuned to maximize either educational value of the content or entertainment value. Or both at the same time. The AI could perhaps ide ntify gaps in your knowledge and adjust the training material appropriately. My mind is blown.

[@Nnm26]

Well if neural link is something to be considered, we may not even need to manually input anything at all.

[@choo_choo_]

Sounds boring.

[@Alex-ck5gf]

@@Nnm26 neuralink is a scam

[@digitalclown2008]

Those stories wouldn't be very interesting as the technology stands right now. AI doesn't have feelings, has no humor, and doesn't feel fear. So, it's only reference for what is "scary" and what is "funny" is whatever the data it was fed and TOLD was scary or funny etc. A visual story on a television screen requires many ideas to be curated and put together to form a cohesive end product. An AI could do this, but it would never experiment. It would never create anything unconventionally new AND cohesive, that would also speak to the human experience.

[@zahimbaqaye3180]

This is the new era for AI. Imagine the reverse role of Dalle 2 where it can render image into text. This would allow even robots to think and Imagine 🤯

[@TracksWithDax]

I reckon that already exists, doesn’t it? I feel like 2MP has covered something like that previously

[@khiemgom]

it alr did, ironically before text to image, someone askes the same question: "Imagine the reverse role of *some image to text engine* where it can render image from text. This would allow even robots to think and imagine *head blows or sth idk emoji*"

[@zahimbaqaye3180]

@@TracksWithDax IDK. maybe he covered this topic also. 🤔

[@IceMetalPunk]

That kind of AI has existed for awhile now. And DeepMind's recent Flamingo AI is able to take a combination of text and images as input (including video, by splitting its frames into an image sequence) and answer questions about the video or image.

[@LutraLovegood]

Once I get my hands on Dall-E I'll try to see how good it is at generating niche stuff, like completely new Medabots, scenes from old video games (Jade Cocoon, Tail Concerto, Dewprism, that kind of things). But I see a lot of potential to create references for scenes that would be complicated to stage irl, like a subway train full of alligators, a naked old man doing acrobatics above a crowd of opossums, Putin eating dirt, exclusive screen shots of Half Life 3...

[@vectoralphaAI]

I wonder what kind of AI we will have 10 to 50 years in the future.

[@mal2ksc]

I wonder what kind of AI we will have 1 to 5 years in the future. Even 10 years looks like it may be beyond the prediction horizon.

[@cappypyramsaudpate5535]

to me its a little frightening i wont lie haha. Show this to a older person who isnt tech literate and they'll be either astounded or worried lol

[@hombacom]

Best AI is the one we cant use yet and we imagine it would solve all our problems

[@Rctdcttecededtef]

If the singularity is anything to base it off of I would say they'd become godlike but tbh that's a random guess. Im not an expert in the field but I think most would agree that's too far off into the future to make an accurate prediction. Ironically the best to answer that question would probably be an AI itself

[@cappypyramsaudpate5535]

@@Rctdcttecededtef ive been thinking about a singularity more often after seeing how quickly AI research is coming along. Every few months were able to do things that were impossible. I just really hope the singularity doesnt torture us forever. We would be at its complete mercy once the exposition happens

[@DonVigaDeFierro]

This does seem like a giant step towards real-time, AI generated VR pornography. I'm still waiting for it.

[@Rctdcttecededtef]

"dall e, generate big tiddy anime girls"

[@billyoung9538]

I'm very excited to try this out my self.

[@xxge]

It can do subsurface scattering but it still made gave the hand 6 fingers! As always there is room for improvement, and I can't wait to see how this field will grow!

[@l.s.11]

I'd love to hear the discussions on how COPYRIGHT plays a role in and alongside this A.I.

[@TheGeekRex]

I've always wondered the same. It doesn't contain any specific images at the end product, but it was trained using literally every image it could get a hold of on the internet. If I typed in all the details of a picture from an artist, it could potentially recreate exactly the original picture. As an artist, what if I don't want my work to be used to train my replacement? What if I don't want someone to be able to type in my name as a style? Do I even get a say?

[@l.s.11]

@@TheGeekRex Fair point, I didn't even think about it from that angle. Reminds me of the whole deepfake and p0rn debate - does one deserve the rights to their figure/presence/image not being used in a transformative way?

[@abderrahimbenmoussa4359]

This will certainly not replace artists but it will for sure multiply the number of concept art they will produce. Mix this with a a tool to generate 3D from images and plug on animator AI and you would be right to expect full AI animation movies !

[@edenkeeping2298]

It will definitely replace artists in a lot of fields

[@Dunixify]

perfect for stock photos /videos for video essays or any script based content

[@Emanresu56]

Awesome, I finally have something that could generate depictions of my video game concepts I randomly daydream about.

[@arbaguincci7591]

It won't democratize art because they won't give it to everyone. They're being really restrictive on who gets to even try it. If anything i'd be worried about the advantages for the people they allow to have it over everyone else. Right now its just being tested but imagine them taking sides in politics or business later on when they decide its ready to be used. They've already shown that its entirely up to them who can create convincing fake photorealistic pictures with the ai, and while the ai is currently curated to keep from making certain kinds of images, there's nothing stopping them from editing it or having a second version for the people who'll pay them enough or for who they prefer, or even just for themselves.

[@beachcomber2008]

"They". YOU.

[@drorange2261]

It will probably be a subscription model a bit like adobe. You need a lot of processing power to create these images, You really need to run a server fully for a bit in order to get good quality results. So it won't be for free, also if they are serious about this they need an amazing dataset , and that will cost thousands of man hours with real world specialists looking at images. What is a mullion, metope, dichroic film ... ? Most photos are not tagged properly even in stock photography websites ! It seems that all of these projects are not interested in allowing the client to do the calculation. To be fair you need many many gigs of ram ... and good gpus to run large dataset ML efficiently. Open AI is for profit corporation with Musk, Thiel and Microsoft on board ... what can go wrong ?

[@beachcomber2008]

@@drorange2261 Stupidity being infinite. . .

[@USBEN.]

That's how it is until community makes it's own version.