Stability AI Launches SDXL 1.0, Its Most Powerful Image Model

Stability AI released SDXL 1.0 today, the latest version of its Stable Diffusion family, the text-to-image generator the company maintains as an open alternative to Midjourney and DALL-E. The company is billing this launch as the biggest quality leap since the project launched in 2022.

What's new

The most visible difference from earlier versions of Stable Diffusion is native resolution: SDXL 1.0 generates images at 1024x1024 pixels from the start, compared to the 512x512 or 768x768 that SD 1.x and 2.x models worked with. That translates into fewer artifacts, better handling of hands and faces, and more coherent compositions without needing to fall back on external upscaling tools.

Stability describes the architecture as a two-stage pipeline: a base model that generates the initial image and a refiner model that adds detail in a second pass. According to the company, the base model has 3.5 billion parameters, and the base-plus-refiner combination totals nearly 6.6 billion parameters when used together. It is, by a wide margin, the largest model Stability has released openly to date.

From preview to public release

SDXL didn't come out of nowhere. Weeks earlier, Stability had distributed SDXL 0.9 as a research preview with restricted access, meant for artists and developers to test the model and report issues ahead of the full launch. SDXL 1.0 is that same architecture, now polished and open to anyone.

Availability and licensing

The model's weights are available on Hugging Face and GitHub, allowing anyone to download it, run it locally, and modify it. Stability distributes it under a license that permits both research and commercial use, a decision consistent with the strategy the company has followed since the first Stable Diffusion: compete on adoption volume rather than tight access control.

Beyond the open model, SDXL 1.0 is already integrated into Stability's own products, including DreamStudio and Clipdrop, and the company has announced availability through Amazon SageMaker for those who prefer to deploy it on cloud infrastructure without managing the hardware themselves.

Why it matters

The AI image-generation market has split into two opposing camps: closed, API-access models like OpenAI's DALL-E or Midjourney via Discord, and open models that anyone can download and tune to their liking. Stable Diffusion has been, since its launch, the centerpiece of the open camp, spawning a huge ecosystem of third-party tools, interfaces like Automatic1111 or ComfyUI, and fine-tuned models built for highly specific styles.

A quality leap like the one SDXL 1.0 promises matters precisely because of that openness: it's not just that Stability is improving its own product, but that an entire community of independent developers and artists gains access to a model with capabilities closer to those of closed systems — without the usage restrictions or per-image costs those services impose.

For companies already building products on earlier versions of Stable Diffusion, the update means weighing whether it's worth migrating their workflows, fine-tuned models, and integrations to the new two-stage format. For end users, the change will show up mainly in images that need less post-processing and in a starting resolution that previously required extra upscaling steps.