Stable Diffusion XL 1.0 (SDXL 1.0) is a high-resolution latent diffusion model for text-to-image synthesis developed by Stability AI. As a successor to the Stable Diffusion v1.5 and v2.1 series, it features a significantly larger parameter count designed to improve image composition, photorealism, and prompt adherence. The model is optimized to generate images at a native resolution of 1024x1024 and supports a variety of aspect ratios without the distortion common in smaller models trained on square datasets.
Architecture and Design
The SDXL 1.0 architecture employs a two-stage generation process consisting of a base model and an optional refiner model. The base model is responsible for the initial latent generation and global structure, while the refiner—an ensemble-of-expert-denoisers—is used to add high-frequency details and improve textures during the final stages of the diffusion process. It utilizes a combination of two text encoders, OpenCLIP ViT-bigG/14 and CLIP ViT-L/14, which enhances its ability to understand complex semantic instructions and natural language prompts.
Key Capabilities
One of the primary improvements in SDXL 1.0 is its ability to render legible text and accurate human anatomy, such as hands and faces, which were frequent points of failure in previous iterations. The model demonstrates high contrast, realistic lighting, and deep color saturation out of the box. Official prompting guidance suggests that SDXL 1.0 is more responsive to simple, descriptive language and requires less "prompt engineering" jargon than earlier models to achieve high-quality artistic or realistic results.