Stable Diffusion 2.1 shipped with a fix for hands and a nudge toward negative prompts

If you made AI images in late 2022, you know the hands. Six fingers, fused knuckles, a thumb growing out of a wrist. On December 7, 2022, a week into the ChatGPT frenzy, Stability AI put out Stable Diffusion 2.1 and listed "improved anatomy and hands" among the changes. The text chatbot had the headlines. The image people got a quieter, practical update.

What actually changed

Version 2.0 had leaned hard on a filter for adult content, and according to Stability's own post that filter also threw out a lot of ordinary pictures of people. Fewer people in the training data meant worse people in the pictures. For 2.1, the team fine-tuned 2.0 with a less aggressive version of the filter. It still removed adult material, with fewer false positives.

The release notes also said 2.1 brought back compatibility with many older prompts, handled art styles better than 2.0, and could render non-standard resolutions, including very wide aspect ratios. The model shipped as v2.1-768, with weights on Hugging Face and code on GitHub, and it was available in Stability's DreamStudio app the same day.

That's the boring part, and I'd argue it's the important part. Most model updates that matter to regular users are like this: a data decision fixes a visible flaw.

The part worth keeping: say what you don't want

Buried in the announcement was a how-to. In DreamStudio you could add a negative prompt by typing a pipe after your prompt, then the thing to avoid, then a weight of -1.0. Stability said negative prompts help strip out problems like mangled hands, extra fingers and blurriness, and that they worked even better in 2.1 than in 2.0.

I think this is one of the most underrated ideas in prompting, and it travels well outside image tools. Most of us only tell the model what we want. But a lot of bad output is a predictable failure (the same stiff phrase, the same cliché, the same filler) and naming it is often faster than describing perfection.

A small warning, though. Negative instructions work best when they're specific. "Don't make it bad" does nothing. "No extra fingers" or "no sentences that open with a speech about how fast the world is changing" gives the model something concrete to steer away from.

Same day, a search experiment

One more thing from the same date. Perplexity, the answer engine that summarizes web sources with citations, dates its launch to December 7, 2022, according to its investor IVP. I can't find a reliable record of exactly what the first version looked like, so I'll leave it there. It's a neat coincidence, though: one product teaching people to name what to remove, another teaching them to ask where an answer came from.