Stable Diffusion XL
By Stability AI
Stable Diffusion XL (SDXL) is a text-to-image diffusion model released by Stability AI in 2023, a larger and higher-fidelity successor to the original Stable Diffusion models, producing higher-resolution, more detailed, and more coherent…
Definition
Stable Diffusion XL (SDXL) is a text-to-image diffusion model released by Stability AI in 2023, a larger and higher-fidelity successor to the original Stable Diffusion models, producing higher-resolution, more detailed, and more coherent images from natural-language prompts.
Overview
SDXL improves on earlier Stable Diffusion releases primarily through scale and architectural refinement rather than a wholesale redesign. Its base UNet — the neural network that iteratively denoises an image during generation — is roughly three times larger than the one used in Stable Diffusion 1.5/2.x, and SDXL uses two text encoders together (OpenCLIP-ViT/G and CLIP ViT-L) to interpret prompts, giving it a richer understanding of language and improved prompt adherence. SDXL introduced a two-stage generation pipeline: a base model generates the initial image, which can optionally be passed to a separate refiner model that adds fine detail in a final denoising pass, improving textures, faces, and small details that earlier Stable Diffusion versions often struggled with. SDXL was also trained natively at higher resolutions (1024x1024 and various aspect ratios) rather than being scaled up after training at lower resolution, which contributed to better composition and fewer artifacts in generated images. Like its predecessors, SDXL's weights were released publicly (under Stability AI's community license terms), making it widely adopted for local and self-hosted image generation, fine-tuning into custom styles via techniques like LoRA, and integration into open-source tools such as ComfyUI and Automatic1111. This openness distinguished it from closed, API-only competitors such as Midjourney and OpenAI's DALL-E, and made SDXL a common foundation for research and commercial image-generation products before newer architectures like FLUX.1 emerged.
Key Features
- Roughly 3x larger base UNet than earlier Stable Diffusion versions
- Dual text encoder setup (OpenCLIP-ViT/G and CLIP ViT-L) for richer prompt understanding
- Two-stage base + refiner pipeline for improved fine detail
- Trained natively at 1024x1024 and multiple aspect ratios
- Publicly released weights, widely adopted for local and self-hosted generation
- Broad ecosystem support for fine-tuning via LoRA and custom checkpoints