In the thriving world of AI image generators, patience is apparently an elusive virtue. Eager enthusiasts of Stable Diffusion—arguably the most popular open-source image generator online—are bypassing the wait for the official release of its latest version, Stable Diffusion XL v0.9, which leaked online Thursday.
The version, intended for cloud testing and not yet destined for public release by creator Stability AI, found its way into the public realm and started spreading rapidly. Joe Penna, a member of creator Stability AI, offered a mixed sentiment.
“Am I proud of y’all, or… opposite of proud?” he mused on a subreddit thread centered on the leak.
Meanwhile, his Stability AI colleague Alex Goodwin confided on Reddit that the team had been keen to implement a model that could run on A1111—a fan-favorite GUI among Stable Diffusion users—before the launch. That plan, it appears, will now have to be hastened.
“We were hoping to, y’know, have time to implement things before launch,” Goodwin wrote, “but [I] guess it’s gonna have to be rushed now.”
Stable Diffusion XL represents an apex in the evolution of open-source image generators. It is designed to compete with its predecessors and counterparts, including the famed MidJourney.
With its ability to generate images that echo MidJourney’s quality, the new Stable Diffusion release has quickly carved a niche for itself. It retains its open-source ethos, offering benefits like compatibility with plugins, privacy assurance, and hundreds of different models.
Such features can entice anyone seeking high-quality graphics without shelling out close to $100 a year for MidJourney. It also requires less complex prompts in order to provide quality results.
However, there is a caveat to the clandestinely acquired Stable Diffusion XL v0.9. The torrent consumes a mammoth 91.3 GB of space, although having the base model and refiner should suffice for operations. Furthermore, it doesn’t play well with prevalent interfaces, and the best option to test it currently is through Comfy UI, a lightweight yet less intuitive, node-based interface.
Stability AI, for its part, is playing catch-up. In a quest to regain control, the company has shut down a version of Stable Diffusion XL v0.9 on AI community Huggingface. Unfortunately for Stability AI, the model has already been decentralized via a torrent, and thus it will be difficult to stop the spread.
This new model comes with a couple of drawbacks. Many plugins integral to the success of Stable Diffusion 1.5—such as LoRAs, Lycoris, Lohas, Controlnet, and others—appear to be incompatible with this model. Additionally, once an image is generated by the base model, it necessitates a refining process for the optimal final image.
Yet, these glitches haven’t dampened the enthusiasm. The model’s visual quality—trained at 1024×1024 resolution compared to version 1.5’s 512×512—and the aesthetic quality of the images generated by the XL model are already yielding ecstatic responses from users.
In such a vibrant community, it’s only a matter of time before all eyes turn towards this new epoch in generative art—and it appears that the chaos around the leak has inadvertently stirred a creative frenzy amongst users.
