SDXL-Turbo

A distilled version of Stable Diffusion XL, SDXL-Turbo from Stability AI generates images in one to four sampling steps instead of the dozens the base model needs. The speedup comes from Adversarial Diffusion Distillation, a training method that combines score distillation from an SDXL teacher with an adversarial loss so quality holds up at very low step counts, enabling real-time and interactive generation on a consumer GPU. Inference differs from standard diffusion pipelines: guidance scale is set to zero and negative prompts are not used, and the model performs best at 512 by 512 pixels. Documented limitations include imperfect photorealism, an inability to render legible text, and weak face generation. The weights, distilled from SDXL 1.0, are published on Hugging Face as a research artifact under Stability AI's non-commercial community license, with commercial use requiring a Stability AI membership. It is commonly used for live preview interfaces, latency-sensitive demos, and rapid prompt iteration before switching to a slower, higher-quality model.