Stable Diffusion XL

A two-stage image generation system with ~6.6B total parameters that produces higher-quality 1024×1024 images:

• Base model (3.5B parameters): Creates the initial image using dual text encoders for better prompt understanding

• Refiner model (3.1B parameters): Enhances the fine details and overall sharpness of the base output

SDXL supports multiple image aspect ratios and uses improved training techniques on a dataset of ~1B filtered images. Released July 2023, it represented a significant quality leap over SD 1.5.