SD-Turbo

Stability AI produced SD-Turbo by distilling Stable Diffusion 2.1 with Adversarial Diffusion Distillation, a training recipe that combines guidance from the original model as a teacher with a discriminator, so the student can produce a usable 512 by 512 image in one denoising step, or up to four, instead of a long sampling schedule. That speed suits interactive tools and modest GPUs. The weights sit on Hugging Face in safetensors format with fp16 variants and load straight into the diffusers AutoPipeline classes for text-to-image and image-to-image; set the guidance scale to zero and skip negative prompts, since the model uses neither. Expect imperfect photorealism, no legible text, and unreliable faces, and reach for SDXL-Turbo, which Stability itself suggests, when quality matters more than speed. Terms follow the Stability AI Community License: free for research and personal use, and free commercial use after registration for organizations under 1 million US dollars in yearly revenue, with an enterprise license needed above that. The repository has not changed since mid-2024, so treat it as a finished research release.