Loading video...
Video Failed to Load
I dusted of disco diffusion yesterday and did some speed optimizations. It's roughly 2x faster as a result of that (not sure exactly..), but it still takes a lot of computation to look this incoherent.
30,123 views • 4 months ago •via X (Twitter)
40 Comments

Love it. I'm still in awe of how this era's imagery were mostly blobs that trick the brain into thinking it's something, but it's really nothing, but your brain keeps saying it is something.

losing and catching things without ever having enough time to settle is really mesmerizing

we really need to revive disco with all the optimizations; I have also wondered if it'd be possible to run different models with the disco framework sometimes I still spin up 5.2 on visions of chaos (really can't be bothered running the notebook versions) 5.2 was the only one that seemed to keep coherent animations without going into purple or red hell vqgan is also still plenty fun new models feel like toys compared to those older ones, I mean significantly more powerful/coherent and whatever but idk there was something magic about those early models

Different models with Disco framework is basically @deforum_art It's been the successor for that since the Stable Diffusion release

@deforum_art yeah I haven't kept up with deforum since probably sd1.5, I've seen some people doing disco style animations with flux even

early gen video just hit different. everything was abstract, weird, and kinda broken—but it had soul. today's outputs are technically insane, but i still miss the REM state of a machine

the best thing about our dreams is how bad we are at generating them

🤌🤌🤌

Making me nostalgic for 2022

Remind me the first public preview you shared on X back then in 1017

@Takyon yeah. Posted a test render that accidentally failed to enable the feature that I was trying to test, and it ended up with 143k likes 😅 (down to 121k after account deletions, etc.). X is weird.

That was the good time

2x faster and it’s still a compute hog

A few nights ago, I created a 2m swampy synth track... and after rewatching your video a few times this afternoon, I realized why I was having deja vu... Respectfully, I merged them...👀

This takes me back

niccce

Does the fork with your speed optimizations is available somewhere ? Would love to dive back to disco

Not yet. I should be able to do that. It's still written in a way that can be converted to the Colab notebook.. but I don't think many people are using that any more (and Colab doesn't benefit from the optimizations unless you're lucky and get an A100 or something)

Still it remains one of the best AI animated experience ever!

The GOAT

Absolutely love this .

I really love the early day diffusions, its an art style in itself 😍

I love this. And now I’ve unlocked a new feeling “AI early days nostalgia”. If anyone have any idea how to revive Pytti I would appreciate the tip very much.

I miss these days

What is disco diffusion? Is it an early video generation model or something? Or an early live video generation model, I should say? @grok

It’d not incoherent, it’s beauty!

🙏 are you sure? I worked very hard on this

Quick question: if someone with a software engineering background (C++ networking stuff) with a vague understanding of ML principles wanted to create something like this where would they start? Any book recommendations? Papers? Thanks! Love your work.

Creepy

yea the old stuff is still slow eh?

how much computation

Each frame was 1280x768 (higher res than most people have seen in DD animation before).. and took about 25s per frame with fairly high settings. I stopped and resumed a bit.. but the whole thing took around 11 hours to render.

@glark0x I feel like we could get this rendering in real time.

@glark0x I just need to make it about 200x faster (at this res). but actually, yes

@glark0x Maybe you could train an Ai to procdurally generate it. Look how fast the LLMs can generate text and speech. Video is at our fingertips!

@glark0x To keep some of the same chaos, I'd probably want to keep some CLIP guidance which is slow. However, an image model could be finetuned to approximate that a bit.. and then use less frequent CLIP guidance on that during gen of each frame. Also, a latent CLIP could be great.

@glark0x Actually, CLIP guidance could be applied to frames of video during generation by a video model too. It's madness due to the speed difference.. but when madness is the goal, then why not.

This is really wonderful work

This and deforum will always have a special place in my heart.

The optics of this at the start were defining ones but it receded under the avalanche of tighter, more literal slop imagery... but the potential for this infinite zoom to become a narrative convention is like the early dormant pianoforte potential of the piano awaiting usage!
