The AI art community will have to contain its excitement a little longer. Stability AI, the maker of Stable Diffusion—the most popular open-source AI image generator—has announced a late delay to the launch of the much-anticipated Stable Diffusion XL (SDXL) version 1.0, which was supposed to be released today.
The late-stage decision to push back the launch "for a week or so," disclosed by Stability AI’s Joe Penna, is causing a stir across the AI community.
"We have a few 'late bloomer' fine-tuned models that are internally blowing us away, but need more time to get a clear sense on which is better," Penna explained in a now-deleted Discord post that Decrypt viewed.
He added that each model carries "pretty big technical repercussions" on community fine-tuning and inference. As the AI art community awaits, Penna said that Stability AI is juggling three internal codenames: Vanilla cream, Zi señor, and Milky weights.
What this means, for those not familiar with the scene, is basically that Stability AI found new models that provide great results, but each one interacts differently with LoRAs (Low-Rank Adaptations) and other community tools.
A LoRA is a file that totally changes the way Stable Diffusion renders an image, making it behave more in the way the artist is expecting. Unlike MidJourney and Dall-E, Stable Diffusion lets people train LoRAs to make the model create images of people, styles, environments, clothing, levels of detail, etc, and these mods are what make Stable Diffusion stand against its competitors.
Penna later clarified that the delay announcement was particularly pertinent to a Discord server of fine-tuners who had LoRAs ready to release. "This message was meant only for them. We had previously assumed interoperability with 0.9," he shared.
If Stability AI releases a model that is not compatible with some LoRAs that have been already trained (even if it was using the leaked v0.9), then the learning curve and the adoption of the new model may suffer—and users would suffer the consequences too.
For context, Stable Diffusion v1.5 is still widely used as opposed to the less popular Stable Diffusion v2.0, because the newer version was harder to use and modify.
Meanwhile, the unreleased SDXL v0.9 has already made quite a splash amongst AI art enthusiasts. The v0.9 version, initially intended for cloud testing and not yet destined for public release, leaked earlier this month and found its way into the eager hands of the public.
"Am I proud of y'all, or... opposite of proud?" Penna mused on a Reddit thread centered on the leak.
The burgeoning world of AI image generators leaves little room for patience, and SDXL v0.9 has been no exception. Its premature distribution suggests a high degree of anticipation for the official release. Still, the abrupt need to adjust plans has left the team to, as colleague Alex Goodwin expressed, "implement things before launch in a rush."
Whether seen as a setback or a suspenseful prolongation, the delay has certainly kept Stability AI in the spotlight. Amid the ripples caused by the delayed launch and the surprise leak, the world eagerly awaits the final product that may reshape the landscape of AI image generation.