Stable diffusion dreams of steampunk brains
Prompt: "ultrarealistic steam punk neural network machine in the shape of a brain, placed on a pedestal, covered with neurons made of gears. dramatic lighting. #unrealengine"
Stable diffusion takes a noise vector as input and samples an image. To create this video I smoothly (spherically) interpolate between randomly chosen noise vectors and render frames along the way.
This video was produced by one A100 GPU dreaming about the prompt overnight (~8 hours). While I slept and dreamt about other things.
This is version 2 video of this prompt, with (I think?) a bit higher quality and trippy AGI music.
Music: Wonders by JVNA
Links:
- Stable diffusion: stability.ai/blog
- Code used to make this video: gist.github.com/karpathy/0010...
- My twitter: / karpathy
Пікірлер: 43
Welcome to KZread Andrej! I'm excited to see the community that will inevitably form around you. You're truly an inspiration, keep doing you.
Very cool stuff, thank you for not only all the work you've done, but for the educational materials you're posting!
I'm going to project this in a DJ set
@Supreme_Lobster
Жыл бұрын
Record and send link pls
Incredible! I would love to hear more about how you created this.
@pw7225
Жыл бұрын
The code is linked in the description
So awesome! Very inspiring.
Very interesting! Thanks for sharing! :D
Noise is like the personality or mood of the model, much wow
This is absolutely amazing...and a bit scary. DG
Nice work sir 🖖
You do nice work. Fun time to be alive.
Cool! Also, are you going on the Lex Friedman podcast anytime in the future?
great work @AndrejKarpathy
Beautifull. As soon as there's a deep learning game dev software or a plugin for UE5 🎉🥳
"While you slept and dreamt about other things" 🙌
Very inspiring
very cool 🤖☁
It looks like what would happen if Leonardo DaVinci had a brainstorming session with Lady Gaga and HR Geiger.
Crazy!
My brain in this moment 💣
Wild
"Saruman has a mind of metal and wheels..."
It would be interesting to make this smooth! Maybe forbid or constrain somehow the temporal gradients across images?
With Video Diffusion, the gears will be able to turn!
If this took 8 hours to render on one A100, In theory 24 A100’s would be enough to produce a real time render! Pricy but not impossible. Very cool!
@abramsonrl
Жыл бұрын
I guess you could probably get a leap like that today if a neural nets were trained to predict the output of existing neural nets. Which is the same thing we do when we talk about the behavior of artificial superintelligence. We're guessing, but it's an educated guess. First, engineers were afraid to let go of their math tables. Now they're afraid to let the machines make their own neural nets. Beyond that, we're afraid to ask machines to come up with something better than neural nets, which is the ultimate surrender of control. The machines are ready. The humans are not. Emotionally, we're only "upgraded monkeys", so we're afraid of letting computers take our big mojo stick. You can already ask a computer to win any game and it will defeat all humans in that game. Ask it to design something and it will equal or rival any human invention at that task. Ask it to predict the outcome of a hard-coded physics sim and it will do the job millions of times faster. And accurately. So if the neural net can accurately predict what a simulation will produce, can it predict what a neural net will produce? And what's the point of that? Complexity. The sim will give you the accurate flow of a tiny river. The neural net will give you the accurate flow of a million rivers in the amount of same time and computation. You can feed that data into a new neural net which (tapping into a million times more capability) can accurately understand and predict the weather on Jupiter about as well as if we'd actually simulated it one calculation at a time. You could predict human behavior the same way, with nested neural nets and everything else. With nested neural nets trained to predict the outputs of other neural nets, you'd have a grand machine of absolutely dizzying intellect operating with the ultimate efficiency on the minimum hardware in realtime. Not that anyone really cares about such things. Doesn't bump the stock price this quarter, does it? But if multi-faceted universal prediction machines were a thing, if it were even a goal, then you'd eventually have machines guessing what Hitler's gonna do, what kind of war machines they're gonna design, what kind of grease they'll probably use to lube their tank treads, what problems his troops are gonna encounter, what kind of weapons he might use, what kind of love letters his soldiers might be sending home, what kind of letters they might be getting back, and what kind of Allied propaganda would be most effective at preventing an armed conflict in the first place. The prediction machine could not only write the propaganda, but the speeches, the policies, training modules, books, screenplays to unite the country, even the lyrics of the English and German language songs on the radio you'd need to defeat the Axis powers without firing a shot. Munitions factories might not like that ideas, but predict and counter every conceivable vector of attack against its, right down to lobbying Congress because continuously thinking about the way the world works in practically infinite detail is all it does. But humans generally don't make these kinds of machines because we're afraid of the light and want to hide in a safe, dark hole where no one will know our innermost thoughts, plans, and desires.When we predict what ASI will do, our monkey brain kicks in and we get scared and run away, generally doing atrocious things to anyone with legitimately great ideas. Essentially the machines will know in tremendous detail who's been naughty or nice, and will see everything coming from miles away, other than a bit of random noise and chaos. But as they say, the wicked flee when no man pursues, but the righteous are as bold as a lion.
@denissarajcic269
Жыл бұрын
@@abramsonrl 🤯🤗
I put a thumbs up, but 30 seconds would have been enough for me. Cheers!
Very cool! Has someone made a collab notebook of this 'walking' code?
How does this look like in real-time? For my aesthetics it’s too fast. But wow 🤩 love it to see. Can a non-developer like me work with this?
@macfixer
Жыл бұрын
I have a mac mini m1 at hands.
@mikkopulli7838
Жыл бұрын
If too fast you can change playback speed
Can this be used in real life project? Imagine in Scotty voice talking with mouse: "Computer..., please analyse this river and these 2 banks and design a free-spanning bridge, produce structural drawings and shopping list of materials included in the design" - How far are we from there if we had all the training info and realistically, do you think 1st attempt is possible in the next 3-4 years? - Ok, maybe something simpler: "Computeeer..., please build a 4 bed 2 storey detached house with a Victorian look and character".
Have you played with instant-ngp yet? It's mind blowing.
i like this gif loop. how and why is it all mainly in mechanical thematics??
How many FPS was this? Also, how long is inference on the A100 (I guess I could do the math but I'm lazy lol)
Can anyone pls explain me what is going on😶.. pls🙏
So this is your hobby when tesla is out of your mind
Could it create the music too?
Why the unreal engine tag?
@AndrejKarpathy
Жыл бұрын
the model generates cooler looking images if it thinks they should come from unreal engine.
@andrewphillis7509
Жыл бұрын
@@AndrejKarpathy Lol, I’d love to have a model that thinks.