SDXL Lightning is a text-to-image generation model developed by ByteDance that significantly accelerates the inference process of the Stable Diffusion XL (SDXL) architecture. It utilizes a progressive adversarial diffusion distillation technique to achieve high-quality image synthesis in a fraction of the time required by standard models. While the base SDXL model typically requires 20 to 50 inference steps, SDXL Lightning can produce high-resolution (1024x1024) images in as few as 1, 2, 4, or 8 steps.\n\nThe model is released in multiple formats, including full UNet weights and Low-Rank Adaptation (LoRA) modules. The 1-step and 2-step versions are designed for extreme speed and low-latency applications, whereas the 4-step and 8-step variants balance generation speed with superior image fidelity and prompt adherence. Because it retains the underlying 2.6 billion parameter structure of SDXL, the model is fully compatible with existing ecosystem tools such as ControlNet and custom LoRAs.\n\nIn its distillation process, the model treats the reduction of sampling steps as an adversarial learning task, training a student model to match the output distribution of a multi-step teacher model. This approach minimizes the artifacts and mode collapse often associated with traditional distillation methods. For optimal performance, it is recommended to use the model without classifier-free guidance (setting the guidance scale to 1.0) and to utilize the specific checkpoint or LoRA weight corresponding to the intended number of inference steps.