SDXL-Lightning

SDXL-Lightning compresses Stable Diffusion XL inference from dozens of sampling steps down to as few as one. ByteDance trained it with progressive adversarial diffusion distillation, distilling stable-diffusion-xl-base-1.0 into checkpoints that generate 1024x1024 images in 1, 2, 4, or 8 steps. Each step count ships in two forms: full UNet weights, which give the best quality for that configuration, and LoRA adapters that can be applied to other SDXL-derived base models to accelerate them. The one-step checkpoint is explicitly experimental, with the two-step and higher variants recommended for reliable output, and quality generally improves with step count. The models load through the Diffusers library and ComfyUI, and the approach is documented in an accompanying research paper. Distributed on Hugging Face under an OpenRAIL++ license, SDXL-Lightning is used wherever generation speed matters: interactive tools, real-time previews, and high-volume or cost-sensitive inference.