SDXL Turbo
Image synthesis from text prompts, instantly.
About SDXL Turbo
Introducing a cutting-edge text-to-image generation model known as SDXL Turbo. This innovative model utilizes a distillation technique called Adversarial Diffusion Distillation (ADD) to produce high-quality image outputs in just one step. By implementing this real-time model, you can expect a significant reduction in required steps from 50 to just one, resulting in improved efficiency and high sampling fidelity. The distillation technique employed in SDXL Turbo combines adversarial training and score distillation, as outlined in the research paper provided by Stability AI. This unique approach enables the model to generate superior image outputs without common issues such as artifacts and blurriness often associated with other distillation methods. In performance comparisons with various diffusion models, including StyleGAN-T++, OpenMUSE, IF-XL, SDXL, and LCM-XL, SDXL Turbo has consistently outperformed these models in terms of image quality and prompt adherence. Impressively, SDXL Turbo surpasses a 4-step configuration of LCM-XL with just one step and a 50-step configuration of SDXL with only 4 steps. This showcases SDXL Turbo's ability to deliver exceptional image quality while minimizing computational requirements. Furthermore, SDXL Turbo boasts enhanced inference speed, capable of generating a 512x512 image in just 207ms on an A100 GPU. Its seamless integration with Stability AI's image editing platform, Clipdrop, allows users to explore and test the capabilities of this real-time image generation model. Please note that SDXL Turbo is currently intended for non-commercial use only. For inquiries regarding commercial usage, kindly reach out to Stability AI for further details.