Generating images with artificial intelligence (AI) has long been a time-consuming process. However, Stability AI has developed a groundbreaking solution that dramatically speeds up this process. With the introduction of SDXL Turbo, real-time image generation is now accessible to the masses. Through innovative techniques and a reduction in generation steps, Stability AI has achieved substantial acceleration in image creation. This article explores the impact of SDXL Turbo and the underlying technology that enables this remarkable advancement.
Traditionally, generating images with AI required significant time and computational resources. However, SDXL Turbo has changed the game by reducing the generation steps from 50 to just one. This breakthrough leads to faster image generation and a reduced compute load. According to Stability AI, a 512×512 image can now be generated in a mere 207ms on an A100 GPU, representing a significant improvement over previous AI diffusion models. This rapid image generation capability is reminiscent of the predictive typing feature used by search engines and operating systems, but for the visual realm.
The speed enhancement in SDXL Turbo is not attributed to superior hardware but rather to Stability AI’s novel research on Adversarial Diffusion Distillation (ADD). By implementing the “One step Stable Diffusion XL” approach with ADD, Stability AI achieves accelerated image generation without sacrificing diversity. Founder and CEO, Emad Mostaque, describes the approach as “way faster & more variants to come” in a recent post on X.
SDXL Turbo is built upon the foundation of the SDXL base model, which Stability AI introduced earlier this year. The SDXL base model boasts several key innovations, including ControlNets that enhance image composition control. With 3.5 billion parameters, the SDXL base model also provides higher accuracy, thanks to its broader awareness of different concepts. SDXL Turbo leverages the advancements of its predecessor and focuses on optimizing performance while maintaining exceptional image quality.
In the quest for faster generative AI models, one common tradeoff has been sacrificing image quality and accuracy. However, SDXL Turbo manages to minimize this tradeoff, delivering highly detailed results that are only marginally inferior to non-accelerated versions of SDXL. This achievement is impressive considering the accelerated nature of the model. Stability AI’s successful combination of diffusion models and GAN-like speed sets SDXL Turbo apart from its predecessors.
The research report on ADD explains that it aims to merge the superior sample quality of diffusion models with the inherent speed of GANs. Stability AI researchers developed the Adversarial Diffusion Distillation approach to outperform other AI methods for image generation. ADD utilizes a combination of adversarial training and score distillation, leveraging knowledge from a pretrained image diffusion model. By doing so, ADD enables fast sampling, iterative refinement, and the utilization of stable diffusion model pretraining.
Experiments conducted by Stability AI researchers highlight the significant advantages of ADD over traditional GANs and other diffusion distillation methods. ADD excels in 1-4 step image generation scenarios, surpassing its competitors in terms of speed, fidelity, and iterative refinement. While SDXL Turbo is not yet considered ready for commercial use, it is already available in preview on Stability AI’s Clipdrop web service, showcasing its impressive capabilities.
Stability AI’s introduction of SDXL Turbo marks a significant milestone for image generation with AI. By drastically reducing the number of generation steps, SDXL Turbo enables real-time image synthesis that was previously unimaginable. The combination of the SDXL base model and the innovative ADD approach creates a powerful framework for creating high-quality images at an unprecedented pace. As AI technology advances, SDXL Turbo paves the way for further breakthroughs in the field of generative AI.
Leave a Reply