SDXL (Stable Diffusion XL)

What it is

SDXL (Stable Diffusion XL) is a text-to-image diffusion model released by Stability AI under the Stable Diffusion name. Stability AI announced SDXL 1.0 on 26 July 2023, following a limited, research-only release of SDXL 0.9.^[1]

A diffusion model makes an image by starting from noise and denoising it step by step, with a text prompt steering each step; the number of steps is the main lever on how long generation takes.

SDXL is built in two parts: a 3.5B-parameter base model and a 6.6B-parameter ensemble pipeline in which a refiner model adds detail. The base model can also be used on its own.^[1]

Technical design

The SDXL paper, submitted to arXiv on 4 July 2023, describes a UNet backbone three times larger than the one in earlier Stable Diffusion models, plus a second text encoder.^[2] The paper also introduces a refinement model whose purpose is to improve the visual fidelity of samples, and it trains the model on multiple aspect ratios.^[2]

Generation happens natively at 1024×1024, and the model is designed to run on consumer GPUs with 8 GB of VRAM.^[1]

Release, license and access

SDXL 1.0 was released under the CreativeML OpenRAIL++-M license, with weights and code published on GitHub.^[1] At launch it was reachable through Clipdrop, the Stability AI Platform API, Amazon SageMaker, Amazon Bedrock, DreamStudio and the Stable Foundation Discord.^[1]

Stability AI described SDXL 1.0 as "the best open model for image generation", citing preference data, and said it handles hands, text and spatial composition better and needs shorter prompts.^[1] These are the company's own comparisons, published with the release.

SDXL Turbo

On 28 November 2023 Stability AI announced SDXL Turbo, a distilled version of SDXL that produces an image in a single step.^[3] The method behind it, Adversarial Diffusion Distillation (ADD), combines score distillation with an adversarial loss.^[4] Turbo reduces the required step count from 50 to one.^[3]

Stability reported that SDXL Turbo makes a 512×512 image in 207 ms on an A100, and that in blind tests a single Turbo step beat a 4-step LCM-XL.^[3] It was released under a non-commercial research license with a demo on Clipdrop, and Stability stated that Turbo was not yet intended for commercial use.^[3]

Place in the line

Within the Stable Diffusion series, SDXL 1.0 fixed the shape of that generation: a base model plus a separate refiner, native 1024×1024 output, and open weights under an OpenRAIL license.^[1][2] SDXL Turbo, a few months later, pushed in the opposite direction, trading tens of sampling steps for one, and was opened only for research.^[3]