Stability AI releases SDXL 1.0

Background

Stability AI's image models belong to the Stable Diffusion line. SDXL was the next step for the company's flagship open image model. Before the full version, the company first released SDXL 0.9 on a limited basis, for research only. ^[1]

What happened

On 26 July 2023 Stability AI announced SDXL 1.0 as its new flagship open image model. ^[1] The release followed the limited, research-only release of SDXL 0.9. ^[1]

SDXL 1.0 has a base model with 3.5 billion parameters and a base-plus-refiner pipeline with 6.6 billion parameters. In the pipeline, the refiner model adds detail to the base model's output. The model generates natively at 1024x1024 pixels. ^[1]

The base-plus-refiner design splits generation into two passes. The base model first produces a complete image at the target resolution, and the refiner then improves it in a second pass to add fine detail. VRAM is the memory on a graphics card, and its size determines how large a model can be loaded and run on a home computer, so a figure such as 8 GB describes the hardware a user needs. Open weights mean the trained parameters can be downloaded, so anyone can run the model on their own equipment under the license terms.

Key details

The model runs on consumer GPUs with 8 GB of VRAM. It is licensed under CreativeML OpenRAIL++-M, and its weights are on GitHub. ^[1]

At launch it was available through Clipdrop, DreamStudio, the Stability API, Amazon SageMaker and Amazon Bedrock. ^[1]

The company then followed on 28 November 2023 with SDXL Turbo, a one-step distilled version for real-time generation. ^[2]

Impact

SDXL 1.0 showed that a large open image model could run on hardware found in ordinary consumer PCs while being offered through several cloud platforms at once. The later SDXL Turbo showed that the same family could also be distilled for real-time use, where an image appears after a single step rather than many. ^[1][2]