SDXS is a novel diffusion model that significantly reduces model latency through model miniaturization and a reduction in sampling steps. It utilizes knowledge distillation to simplify the U-Net and image decoder architectures and introduces a novel single-step DM training technique employing feature matching and score distillation. The SDXS-512 and SDXS-1024 models achieve inference speeds of approximately 100 FPS and 30 FPS on a single GPU, respectively, which is 30 to 60 times faster than previous models. Furthermore, this training method holds potential applications in image-condition controlled generation, enabling efficient image-to-image translation.