Is Meta's New AI Image Generator Good?
HTML-код
- Опубликовано: 14 авг 2024
- Today we're taking a critical look at Meta's New AI Image Generator, Imagine. Plus, a remarkable new AI Avatar that utilizes Gaussian Splats to create the next level of virtual technology. All that, plus a look at the newest AI Video Model!
🔗 Useful Links:
Meta's Imaginer:imagine.meta.com/
Relightable Gaussian Codec Avatars: shunsukesaito....
WALT: walt-video-dif...
👉 Stay Connected:
Twitter: / theomediaai
Join the Discord: / discord
💡 Your Thoughts:
What do you think about Imagine?
Do you see AI Avatars as the future of virtual meetings? What other use cases do you see for it?
🙏 Support the Channel:If you'd like, please join the Patreon Here: patreon.com/TheoreticallyMedia
Like, Comment, and Subscribe if you enjoyed the video.
Hit the bell icon to stay updated on our latest uploads.
📺 Watch Next:
Prompting a World & Animating Anyone! Plus: Microsoft’s Gaia and More!
• Stunning Breakthroughs...
Remember: This channel is all about exploring the creative potential of AI tools. Join us on this journey to discover how AI is revolutionizing the creative world. Thanks for watching, and I'll see you in the next video!
00:00 🧠 Meta released a free text-to-image model called Imagine, trained on 1.1 billion Instagram and Facebook photos.
00:37 📸 Imagine generates images with decent quality, but some minor issues like compositing.
01:49 Examples from Imagine
02:50 👨👩👧👦 Real Training equals real people?
03:20 📷 Imagine can create images in the style of professional Instagram model shoots.
04:16 💡 Imagine shows creativity with prompt "cinematic kung fu movie animated 3D," despite some flaws.
04:59 🆓 Imagine's real strength is its availability and the fact it's free, offering potential for improvement.
05:27 🤖 Relightable Gaussian Kodak avatars capture 3D details at a submillimeter level and offer dynamic facial expressions.
06:32 🌐 AI avatars in VR meetings are becoming more realistic, with potential for practical applications.
07:15 🎥 The AI video model "Walt" uses an attention module to focus on relevant areas, improving coherency and movement in videos. It's a research model with room for improvement. 
Наука
Hi Tim! I'm with you that AI needs to be able to more easily generate people that portray realistic everyday human beings. I simply wanted to point out that it is possible to obtain them with stereotypically "pretty" models, with just a bit of prompt-fu 😁
For example, with Midjourney I got some nice results using:
"Here's our family photo during our disneyland vacation! Love you auntie! Looking at the camera, close up --v 5.2 --style raw --stylize 0 --seed 1"
and similar prompts.
Keep it up!
Good call
Thanks Tim for this update. Amazing AI tech waiting to be explored. Very promising. Cheers all!
I’m putting together a year end wrap up of AI’s 2023 journey. I can’t wait for you all to see it. When you lay it out, it’s insane.
Looking forward with eager anticipation. Much obliged.@@TheoreticallyMedia
"Old school busted hands" is hilarious to me. You're right, it is old school in AI time, but like that's only 6 months ago that was an issue 😂 This tech is moving at such a crazy speed.
Busted Hands and Crazy Extra Limbs! I gotta admit, I get a little nostalgic when I see them!
7:08 its going to be kind of funny. Like to be professional youre probably going to use the hyper-realistic skin suit avatars with your employers, but once you get "off work" and hang with your buddies you can switch into something more stylized and custom.
Imagine someone coming to work in a furry avatar, looking at his hands noticing they are paws and saying "oh crap" and changing out of it.
Hahaha-- The future is ripe with so many comedic ideas! But that is a really good point, we'll probably have different avatars for different social situations...wild times ahead!
Great video as always Tim
Thank you so much!!! This was a bit of a weird one, but man- those AI Avatars are CRAZY.
I didn't particularly care about a new AI but I liked the diversity and the normal normal people not the supermodel normal of most other AI. But you got the like from me and my first comment because of the way you say coffeee. Mmm coffeee! ☕
Haha, the magic of black liquid gold! I raise my steaming mug in greeting to you!
I'm learning English through your videos :D, nice video.
Oh man, really? That's kind of awesome. Totally honored! Also, if you get corrected at any point, please consider them right!
Imagine on Meta not available everywhere. I'm in Australasia, and it says it's not yet available in my region...
Sorry to hear. I did mention in the video, it’s stupidly only available in the US right now. Hopefully that doesn’t last long
I'm going to be crass - but realistic here -- in saying that if you want images of pretty young girls, Meta's model will probably deliver extremely well given the training source. And this brings up the interesting thought which is how much a training source will bias a model.
oh, I'm with you 100% on that. It's funny, I think earlier models (MJ v4, maybe early Leo) would have been primarily trained on higher end fashion photography. IG...well-- I'm not saying there aren't very talented model photographers on it-- but, there are a lot of beginners as well. I'm almost curious to see if Meta's model will reflect that.
Little girls are now growing up with a soul crushing experience. Not only are supermodels much, much prettier than they are but AI generated avatars are much, much prettier than even the prettiest supermodels. Somebody needs to increase the college output of psychologists specialized on body image personality disorder... there is an epidemic of it coming our way.
The generated cp and deepfakes are more concerning imo. We at the point anyone from school could take a photo off media or sneak a photo of u at school and now they got u naked or whatever, and yeah, Photoshop has existed, but we’ve seen how good AI at in painting and out painting especially with like adobe firefly, AI has a magic accuracy to it when it comes to guessing what should be there
@@kinkanman2134 We have a very well known way to take care of that problem. The people in the old country used to call it "the law" and "sending someone to jail", which was a pretty horrible house with a lot of people who had done similarly bad things. :-)
No seriously. There is no technological solution to this. Technology is a tool, no different from a hammer. One can build a house with it or... you know... use it to hurt people. The law doesn't try to regulate hammers. It regulates the behavior of people. So, yeah... if somebody circulates an AI generated nude image of somebody else without their explicit permission and that gets them three years without parole, then most people will simply abstain from that nastiness on the grounds that three years in the big house are just not worth the ten seconds of snickering by some other jerks.
Now, should we regulate the generation of such images for ones own consumption? This is where it gets tricky. We can, after all, also not regulate somebody's imagination. If I want to undress somebody mentally I can. There is no law against that and there shouldn't be, unless we all want to live in Orwell's mind crime society. There is, of course, a limit of what my imagination can do and it is, at least currently, completely safe against abuse by others. Nothing "leaks out", so to speak. Achieving a similar level of privacy with digital media is, at least currently, virtually impossible. So, where do we want to set the line? If thinking nasty stuff isn't punishable, should computer aided generation of nasty thoughts be?
@@kinkanman2134 Oops... looks like RUclips doesn't like my flowery language describing the legal angle of the problem.
To make it short: we don't regulate tools, we regulate intent and damages by law. If somebody does nasty things with technology, they will have to accept being fined and jailed for it. If the abuse of these technologies becomes rampant, then we have to increase the level of penalties.
What we can't regulate are mind crimes. If somebody wants to think something nasty about somebody else, then that's well within the privacy rights of the "perpetrator". Should we make computer aided generation of nasty thoughts a mind crime if these images stay private? That's a fine line. Better think about the consequences of erring on the wrong side of it.
@@lepidoptera9337You're right and yet... I wonder if it's possible to find who's doing those things. The race to make better AIs is all about speed and money. No one seems to want to consider including ways to track stuff that's made with their AIs for fear of wasting time and losing competitive edge. Fines and jail time for bad users are useless if those bad users can't be identified. I'm not even talking about it being too late because the harm would've been done because we don't punish thought crimes, as you said.
@@Madelyn24 Yes, it's absolutely possible... if you don't mind Chinese style spyware on your computer and phone. It gets a lot more complicated within a free society because now we need cause and a judge to sign search warrants. One can, of course, restrict AI providers by criminal liability, which will probably collapse the generative AI field quite quickly. Capitalists don't like legal liability. It's incredibly costly because the lawyers take an arm and a leg.
Maybe the European Union will go that way. They are already indicating that they want to act on it if I am not mistaken.
We may have a testcase for this in child pornography. I don't know how effective our current law enforcement is, but it's probably worth doing some research. I would expect similar success and failure rates for other kinds of privacy incursions.
I do agree, by the way, that attacking somebody this way is a heinous act that has to be punished effectively.
I've never been this early to a vid of yours TM, noice!
oh, that's awesome! Early bird Award!!
Nice one Tim
How does it work if you ask for Non-American looking people though . Europeans, British etc
interesting question, I'll give it a shot! Technically, since it's facebook/meta, I think it should know...for better or worse, it is a global platform.
It's funny, I was watching the Beckham documentary on Netflix and I forgot the whole thing where The Sun launched a world wide campaign to find someone in the world who did not know who David Beckham was. They finally found one farmer in Chad...
In a weird kind of way: That's pretty much the reach of Facebook.
I know in MJ and the 'like'. if you ask for anything like a 'British' police officer you get an American 'Cop' Same with Firefighter or Farmer, Miners etc. It seems to default to American images. Even worse if you ask for regional stuff like Welsh, Scottish or Irish . You get the 'American' version of them :) @@TheoreticallyMedia
👋
W.A.L.T. seems to have promise. I think they have a good concept with focusing the fidelity on the most important part of a motion clip.
Hi Louis!! Yeah, it’s a smart idea. They kind of made mention in the paper (or something I read) that the AI focuses as the human brain does. Like, when we see a rabbit running in a field, we aren’t paying attention to the trees swaying in the wind.
I thought it was a pretty interesting idea to incorporate!
@TheoreticallyMedia
Yes! You're slowly compelling me to start to take video AI more seriously than I have previously.......but it was for good reason I didn't...... the quality was and still largely is awful.
But I'm eagerly awaiting the day when they (companies and individual developers) break that fidelity barrier in a legit way. 👍
Tim can I use the images create by Imagine Meta for commercial purposes? Is there a way to find that info?
ooooh, good question I have not looked into the T&C there-- interestingly, after digging around, I can't seem to find anything that says one way or the other. The Terms on the Imagine site take you to Facebook's general T&C...so...yeah, I'm really not sure.
@@TheoreticallyMedia yes I tried to look at the T&C and it behaved like you mentioned.
I guess they will eventually get to making a more reasonable site with more info on Imagine and a relevant T&C
Awesome video!! OMG when are they going to lose the absolutely creativity-killing FILTER CENSORSHIP?!? It’s like they have absolutely no imagination…ironically since it’s called “imagine” Maybe I’m the only one who notices because I am 99th percentile on “openness” or creativity. I tend to see way beyond other people’s interpretations on things but seriously I don’t understand how they are getting away with censoring generative tools for art. It’s completely insane from my point of view and I don’t even like or want to make anything x rated at ALL. It’s blocking almost everything I want to do at some point-by extreme insane prejudice just like book-burning is. First company makes a tool that doesn’t throw out my work or ideas and I will drop the others like a red-hot-nightmare… I can’t be the only one. I can’t even get into “flow” state because I’m too afraid of the storm-troop censorship. Freedom is NOT a dead idea. It’s MUCH MORE IMPORTANT in an AI world. MUCH MORE!!! ❤
Why isn’t everyone talking about this? It was already a problem with search engines… and forums like Facebook. More and more the “real estate” of our lives is online. Our nation is moving online. We should bring the bill of rights with us.
Oh, look-- I'm with you. Obviously there are certain keywords that need to be blacklisted, which it doesn't take too much imagination to recognize. But, I was just saying in another comment, that Meta's Imagine has a pretty heavy hand.
I think it'll lighten up as time goes on. The language models will get better, moderation tools will get better-- and larger than that, the fear of ending up front page news will alleviate as we become more comfortable with these tools.
I can almost guarantee you that someone from Meta's PR and Legal teams are literally sweating in bed every night, worried about a news story breaking that an 10 year old girl generated up some horrific image while trying to make a My Little Pony image.
Sometimes you can get a result by tricking the model as well-- I once needed a image of a bunch of dead zombies in the street (think "The Walking Dead") and Midjourney steadfastly refused to give it to me. But oddly, when I asked for "Zombies taking a nap in the road..." well, I got it!
They aren't dead...they're just REALLY tired!
I'm an artist who is also a grandfather and I have children in the family interested in AI generated art. There are several sites I don't want them going to. I think you are talking about an 18+ world. Sorry, it doesn't work that way. Some things shouldn't be seen by young children.
@TheoreticallyMedia THANK YOU 🙏🏻 for this absolutely awesome reply!!! You made more sense than anyone else I ever had a reply from. I thought myself it’s due to corporate fears… but I live in silicon valley and the reason people leave older or not so old corporations is due to this same dynamic of fear and stability vs vision beyond what the people at the old company want or are able to envision. So people leave and start their own company. Thats nearly impossible since it takes billions to properly train and develop the AI. My friend Federico Faggin who lives in the hills on the edge of Silicon Valley was the first to make an actual full computer on a chip. He left Intel for these exact reasons. So we probably just have to wait and someone will wake up, 😂 👍🏻
The thing that’s really scary is that we all are forced to use the kids “big crayon” version conceptually. All despots see and treat the people like children. There’s no adult or unrestricted version that people who are extremely creative can use who actually think far beyond the filters. I am a paying user (30/month) of Midjourney but the filters are so insanely bad and they have steadily been getting worse for 6 months that I stopped using it and I didn’t even notice or understand that it was automatically “punishing” me for being freely creative…. And was a negative reinforcement. I started out with literally 20 hour days and tons of complex thoughts and ideas… but half way through a day’s long project of many epic threads and preparation it blocked at the key moments and generations!!! Or somewhere along the way. It’s like investing. Lose a ton of money, and instantly you lose interest in the investment. Also, it kills the flow! The fear of blocking is like being a Rembrandt who has a color removed from the palette randomly in the middle of the process…
So anyway, there should be an absolutely unfiltered version for adults. With a paid subscription gateway so kids literally can’t get in. Then we’re golden. Not sure why there isn’t yet.
Thanks again!!! LOVE your videos!!! They are the best on RUclips!!! ❤️❤️❤️
@TheoCage That’s exactly my point-there are MANY ways to make a gateway or other methods so that kids are excluded from a freely creative version. I would be extremely excited about what people could make never dreamed of before: stuff not possible with other media, and the free flow of images can absolutely heal people if they are allowed to explore them without shame. Thats why art is therapeutic-exploring your own mind and ideas is enlightening. Could make a MUCH better world. I trust people. I don’t think people will want to only make ugly things at all. I think they will gravitate toward the transcendent and beautiful
@@TheoCage that’s the whole point-it’s something that needs separate versions for adults who are not offended and willing to risk the randomness for the gold mine of inspiration, synergy with creativity. Its so
Perfect for that, so resonant with life itself and how it works… it cannot be a coincidence. There’s higher laws at work. In fact, I’m finishing writing a book about exactly that. 🤷♀️🙏🏻😂 Just re-wording and trying again… 😊
From my own experience, it's VERY good at stills. Like a person standing, or a car, or a tree. It's as good or slightly better than DALLE-3 in that regard. It's particularly good with faces - not surprising given it trains from Facebook images. But for 'moving' photos, it's not good at all. Type in "100m dash" and watch the funky things it does with their legs. 😆 Literally half the runners don't have feet. Or they float 3 feet off the ground. I tried the prompt several times with the same results.
It also does weird things with crowds. One or 2 people look great, but if you type in "beach party" or something, some of the background people look like a zombie apocalypse. People with no ears, backwards heads, 3 eyes, no eyes, 3 arms, etc. Scary stuff.
Any links to walt?
No wait list! Just hop on! Although you do have to have a Facebook or Insta account...I mean, it's Meta, so I get it.
Sorry Tim. I meant any links where I can try WALT? Thank you!
As always great stuff Tim! Pretty sure no one is looking at her nose ;)
Thanks for the "heads up". Meta's Imagine looks promising. I wanted to create "An Illustration of Sonny Crockett from Miami Vice wearing sunglasses and holding a handgun". All four of Imagine's output didn't quite look like Don Johnson; but they were close enough to the description anyway.
Adobe Firefly had a problem with the word "handgun" and would not render. When I removed that word, Firefly created four images, but two were women and the men were black. None were white men like Sonny.
Bing was even worse: It rejected both "Sonny Crockett" and "Don Johnson" from the get go, even when "handgun" was removed from the prompt.
If you want a specific ethnicity or look you have to specify it. I have to do that all the time if I want people who look like me, then I have to go further and find prompts or ways to tell it to make it look like the people made an effort to comb or style their hair because it likes to create unkempt, dry cotton ball hair for African American children. Then I have to figure out how to get various different shades of brown. Then I have to figure out what word am I using that is causing it not to generate my requested image. 😅
Meta often rejects any celebrity name, but sometimes if you put “ inspired by “ - character name, maybe movie name, or misspelling the name so it’s still phonetically the same. Ex Rihanna to “ RiRi” and it gave me her face.
@@InspiredByEbonyLove, thank you for your very useful tips!! ❤
Good compared to which universe?
Good question. I suppose against Dall-E, Midjourney etc....although, I don't like doing those "Versus" type things.
Can imagine be run locally? Can it be trained? If not I'm personally not interested in the slightest.
I would use any and every other alternative before I used any Zuck muck.
Totally fair. The dude did really screw us over as a society a few years back. Hold that grudge as long as you’d like!
Can't wait until the avatars are used to make Skyrim NPCs better
Ha! That’s be grand! I can study the pores on Lydia’s skin as she scowls at me for the 4000th time!
the b*ing over training from people is silly. On one end, you have a total lack of understanding on how the AI works - I.e. "Omg the A.I. steals my images and then copy and pastes it into its own image and somehow that how the A.I. is able to make a hippo with purple poka dots!, theft!" (As if that even makes sense, on the face of it. And then on the other end, you have this gripe over A.I. not asking permission to be trained on what it sees, which is ridiculous. Every single human's own intelligence is based on "theft" of visual and audio data that the human brain, much like the A.I. neural network, uses to understand a vast array of concepts. Also, uh yeah, Insta and Meta do own your images you have posted on to the platform - so why make a controversy now. Even prior to this A.I. tech explosion, meta owned you uploaded images and every employee there (not saying they did) could see and analyze your content - and they never have to or have ever asked prior. So it's really a non-issue in my opinion.
Oh, 100% agree-- particularly on the T&C of Facebook and Insta. And yeah, it's not like I can prompt my name and have me (or anyone else) show up. But, in terms of overall esthetics? The fact that its probably got trash Digital Photos from 2007? I mean...for me, that's pretty cool.
We're going to train it based on your family's and friends' photos. Now that we have not just implied, but _stated_ that we _will_ do this, let's not fuss with the asinine complaints later on when we see our partners on a beach with our neighbour, Bruce Lee.
It takes Mafia coconuts to put it in black and white, but there it is. We honestly will have no one to blame, which we like to do.
Thanks for the video Tim. 👍
Oh, US only! Yeah, no one else posts. We're safe!🤣
nice and where are the links :)
Sorry, had to Post and Ghost (dinner time/etc)-- they're up now!
Does Meta's New AI Generator come with legs?
Haha, I actually did a little photoshop gen fill on those legs and man…it was not kind to that poor fake model.
Try prompting zombies, it will not do it.
oh really? I only ran into an issue with Bruce Lee (for the Animated Kung Fu example)-- I presume the name drop was a no go. So, yeah-- likely pretty censored. It is Meta after all.
Is it too much to ask for you to ads the link to META's site in the description box?
He just told you how to get there on the video. Is it too much to ask not to get this lazy?
Yeah, I'll pop it in. This was a Post and Ghost. Had to be on a podcast literally right after I published.
Imagine all the people .... 😅
Oh, that's the blue ribbon comment winner right there!
Mommy what happened to that man's face
Haha, Kid, your Dad has been an AI Avatar this whole time!
Is it open source
Which one? Meta's Imagine? I doubt that-- considering...well, it's Meta. WALT (the video model) is. Unsure about the Avatars-- OH man, I just checked: One of the authors of the Avatar paper: META! Geez, they really are going in full with AI.
and no link? so sad...
Link is up, sorry had to post and dash.
🗿🍷
Thank you!!
It stopped working.
First wooohoooooooo🎉o🎉😂🎉😂🎉😂🎉
Winner of the CHICKEN DINNER!!
I'm so proud to be an American and living in the USA.
I can´t help but this is Dall E 2 at best. Considering the quality of Stable Diffusion and Dall E 3, this is a useless toy. Nothing we haven´t seen before. Except you are on Facebook and prompt yourself, lol
Short answer. No.
Yeah, pretty much. But I do stand by that it is one to watch. It might develop into something interesting at some point, there’s some potential here.
Using people’s personal photos to this without their permission is completely unethical imo
I'm sure they'll add an "Opt Out" at some point-- but, I'll say: It was all in those Terms and Conditions we clicked on when we opened a Facebook or Instagram Account. Not that I don't agree with you-- but I mean...it's Meta? They've done WAY worse stuff than using our images.
That's one that I've kind of given up on-- I mean, when Google releases their Text To Video model, I'm sure "Guy in a Black Tee Shirt, standing at a desk, talking about AI" is going to be in there.
@@TheoreticallyMediastill doesn’t make it right. This directly can end up help make even more deepfake corn.
Meta being meta as usual 🤢🤮
🤭🤭🤭 Bro 🎉🎉Bing Ai way better ..
I do like Dalle as well!
I instantly appreciated the ethnic diversity.
Agreed. I’m always looking for it when using very general terms like “Family” or “woman/man”
Lots of people in the world, and while I get that the generators are only providing 4 images at a time, i mean, they should at least represent 25% of the types of people on our little blue dot.
White people have the most diversity so we need more White people. Whites have blue green hazel brown eyes...everyone else...just brown. Hair color blonde brown black red etc...everyone else....black and dark brown.
Stop trying to exclude White people.
not a good generator
Admittedly, it isn't blowing my mind-- but, I think for a very specific use case of realism in AI images, it has a little extra edge. And not realism in the traditional sense, but more from a "these look like normal people" kind of way.
I mean...It's a weird flex? But-- if you're looking for a specific type of image, it might develop into a "go to"
Thanks for spreading information🤍
always honored to!