Run this yourself: [ Ссылка ]
If you can’t / don’t want to run it locally: [ Ссылка ]

Playing around with the latest version of the “disco diffusion“ AI art generation software. I asked the AI to generate a self portrait, and the results are impressive but mixed. Definitely stick around to the end to see some stills from earlier runs, which I didn't render out all the way into a video.

I’ve been doing some fun live streams on my main channel (Moby Motion) where chat messages are automatically turned into images using VQGAN + CLIP. The results are pretty impressive, but it turns out the state of the art is even more advanced. Fundamentally, this works differently (it uses diffusion, as the name would suggest, rather than being a GAN). But on top of that, this is based on a notebook which has some really nice features. On top of this, the notebook I’m using implements something called “Turbo“ which was designed to improve temporal coherence (and it does), but also vastly speeds up rendering.

Not to mention that much higher resolutions seem possible with disco diffusion compared to the old VQGAN + CLIP. I used to top out at 600 x 600 resolution, anything higher would result in memory errors. On the other hand, this was synthesised using Disco Diffusion at roughly 520p (uploaded at higher resolution to give you the best possible bit rate).

Has been amazing watching the steady improvement in this tech over the past few months, which includes some other models like pytti. Love the other channels making this kind of content (and tutorials), like Artificial Nightmares, Nerdy Rodent, Artificial Images, Jeff Heaton.

Music:
Audial - Summit

свернуть