These Macarons Are Not Real

I’m in the private beta for Stable Diffusion. Even with all the large language models of the past few years, there’s something akin to witchcraft when you’re able to type “a photo of macarons” and a Discord bot returns you images like this:

Stable Diffusion-generated macarons

It’s a large step above the VQGAN methods I was using last year and well up there with the output of DALL-E 2 and Imagen. Stable Diffusion has another advantage - soon it’ll be available for free and runnable on really quite low-powered GPU machines. I’m hearing that there’s even going to be a M1/M2 optimized version that will likely generate images in under a minute. When the weights get released in months? Weeks?, it’s Going To Be A Lot. As ever, 2000AD got there first, long ago, back in the Meg, a dark future where IPC sub-editors could dictate our youth without needing pesky writers or robots:

Kenny Who?

It’s interesting that Stable is looking more threatening than GPT-3 has turned out to be so far. I think that’s down to a few factors - where it and its successors fall down is in that longer context - even PaLM, going out of its way to be as helpful as it can (pretending to be sentient if you ask it to, even!), cannot generate a correct and coherent long-form piece. But you don’t need that with an image, because it’s just immediately all there — sure if you look closely at the macarons you’ll see the ganache line can sometimes get a little wonky, but your mind filters over that at a glance, whereas when you get to the first wrong fact in a GPT-X text, it’s pretty much over.

Additionally, the more impressive text models are really quite expensive to train and run, resulting in access to something on the power of GPT-3 being walled off behind OpenAI’s for-pay API. I think that is going to change in the very near future, though. Tim Dettmers’ work is going to bring some of those larger models to your desktop, and Facebook has already released publics weights for the equivalent of the largest GPT-3 language model. We might see a lot more weirdness and exploration in this area in 2023.

Anyway, Christmas cards this year are going to be fun.