SDXL Turbo: Real-time Text-to-Image Generation
SDXL Turbo achieves unparalleled performance through cutting-edge technology, enabling the creation of high-quality images in a single step, a significant leap from the traditional 50-step process.
Utilizing innovative distillation technology known as Adversarial Diffusion Distillation (ADD), SDXL Turbo undergoes training to emulate its teacher model, SDXL, in a single step. The unique aspect of ADD lies in employing a GAN discriminator to ensure that the student model, SDXL Turbo, produces images of the highest quality, mirroring those generated by the teacher model.
One of the standout features of SDXL Turbo is its substantial enhancement in inference speed. On an A100 platform, SDXL Turbo is capable of generating a 512x512 image in a mere 207 milliseconds.
Derived from the SDXL base model, SDXL Turbo undergoes fine-tuning and is specifically trained for the generation of 512×512 images.
It is crucial to highlight that SDXL Turbo is currently not intended for commercial use.
-
For access to Huggins Faces using SDXL Turbo, visit this link.
-
Explore more about SDXL Turbo and its research at Stability AI’s official announcement.
-
Experience the live implementation of SDXL Turbo at ClipDrop.
-
For a deep dive into the ADD methodology, refer to the adversarial diffusionndistillation Paper.