Jon Stokes runs through the changes that Stability AI have introduced to the new versions of Stable Diffusion.
On November 24, 2022, Stability AI released the 2.0 version of Stable Diffusion. Then just two weeks later, they pushed out version 2.1. The short span of time between 2.0 and 2.1 wasn’t solely because the company is trying to iterate faster. Rather, 2.0 was widely criticized by its large community of users that had embraced the initial release of Stable Diffusion 1.5.
Many of these users, myself included, fiddled around with 2.1 and found that none of our prompts were giving us the result we wanted, then went right back to using Stable Diffusion 1.5. The subsequent 2.1 release tried to address some of the criticisms, but in the view of many, it didn’t go far enough.
In a nutshell, Stable Diffusion 2.0 and 2.1 have the following things working against them, some of which are just the inevitable growing pains of innovation but others of which are the result of the AI wars that we’re just in the opening battles of.
A new text language interface that subtly invalidated most users’ intuitions about how to work with the model.
A totally remixed embedding space that effectively scrapped a massive, crowdsourced map of 1.5’s higher-dimensional latent space.
More guardrails against generating NSFW content.
Read More at Jon Stokes
Read the rest at Jon Stokes