The next version of Stable Diffusion won’t produce spaghetti hands

SDXL 0.9 ‘presents a leap in creative use cases for generative AI imagery,’ said Stability AI.

Left: SDXL Beta, Right: SDXL 0.9 (Stability AI)

The next version of the prompt-based AI image generator, Stable Diffusion, will produce more photorealistic images and be better at making hands. SDXL 0.9, a follow-up to Stable Diffusion XL, “produces massively improved image and composition detail over its predecessor” according to Stability AI. The announcement appeared in a since-deleted blog post and was reported by Bloomberg.

Stability AI describes the new model as providing “a leap in creative use cases for generative AI imagery.” Sample images included in the blog post revealed advancements when using the same prompts in Stable Diffusion XL beta and SDXL 0.9. The images generated with the newer model — including aliens, a wolf and a person holding a coffee cup — appear to show finer detail and more convincing hands. Hands were an easy “tell” to spot AI-generated art — at least until Midjourney v5, a rival platform that runs on Discord, launched in March.

Comparison images of AI-generated artwork, featuring an Alien in Las Vegas.
Stability AI

“Despite its ability to be run on a standard home computer, SDXL 0.9 presents a leap in creative use cases for generative AI imagery,” Stability AI said. “The ability to generate hyper-realistic creations for films, television, music, and instructional videos, as well as offering advancements for design and industrial use, places SDXL at the forefront of real world applications for AI imagery.”

Stability AI writes that the new model’s “significant increase in parameter count (the sum of all the weights and biases in the neural network that the model is trained on)” allows for the improved results. Running SDXL 0.9 locally on a PC will require a minimum of 16GB of RAM and a GeForce RTX 20 (or higher) graphics card with 8GB of VRAM. It supports Windows 11 / 10 and Linux.

Split-screen sample images of an AI-generated photo of a wolf
Left: SDXL Beta, Right: SDXL 0.9 (Stability AI)

According to the deleted blog post, the model will soon be available on Stability AI’s Clipdrop web tool and will be added to the startup’s DreamStudio app. The startup says the open-source version of SDXL 1.0 will arrive in mid-July.