Stable Diffusion XL
Stable Diffusion XL (SDXL) 1.0 is Stability AI's latent diffusion text-to-image model for native 1024x1024 generation.
Provider
Stability AI
Model family
Stable Diffusion
Diffusion image model
Cost tier
Xl
Status
Current
Release Jul 26, 2023
Why teams choose it
The base checkpoint can run standalone; add the refiner only when its measured quality g…
ain justifies the extra inference work.
Safety filters, watermarking, and abuse controls are deployment responsibilities for self-hosted workflows.
Safety filters, watermarking, and abuse controls are deployment responsibilities for self-hosted workflows.
Pin the exact checkpoint, scheduler, inference steps, resolution, and hardware when comparing latency or cost.
Pin the exact checkpoint, scheduler, inference steps, resolution, and hardware when comparing latency or cost.
Tradeoffs to know
- The official model card notes imperfect photorealism, unreliable legible text, compositional errors, and possible face-generation failures.
- The base-and-refiner workflow adds inference work compared with running the base checkpoint alone.
- Track CreativeML Open RAIL++-M license obligations and complete a deployment-specific safety review.
When not to use this
- Self-hosting outcomes depend on hardware, quantization, and ops maturity—budget time beyond swapping an API hostname.
- May demand more instrumentation than SaaS-managed APIs to duplicate latency, failover, and support guarantees.
- Benchmark prompts and regressions continuously before rewriting entire routing tables around weights.
Technical specs
- Inputs
- text, image
- Outputs
- image
- Capabilities
- text-to-image, img2img, base-and-refiner workflow
- License
- CreativeML Open RAIL++-M
- Model string
stable-diffusion-xl
Benchmarks
No benchmark data yet.
What is Stable Diffusion XL?
Stable Diffusion XL (SDXL) 1.0 is an open-weight latent diffusion model for text-to-image generation. Stability AI published a base checkpoint that can run on its own and an optional refiner checkpoint for the final denoising steps. The architecture increases the UNet size relative to earlier Stable Diffusion releases, uses a second text encoder, and adds size and crop conditioning.
Stable Diffusion XL 1.0 specifications
Reference details for the published SDXL 1.0 base checkpoint. Verify the exact checkpoint and inference pipeline before production benchmarking.
| Specification | Value |
|---|---|
| Release date | July 26, 2023 |
| Native image size | 1024 x 1024 |
| Workflow | Run the base checkpoint standalone or use the optional refiner for the final denoising steps. |
| Architecture | Latent diffusion with a larger UNet than earlier Stable Diffusion releases, a second text encoder, and size and crop conditioning. |
| Published weights | SDXL 1.0 base and refiner checkpoints are published by Stability AI on Hugging Face. |
| License | CreativeML Open RAIL++-M |
Stable Diffusion XL vs Stable Diffusion 1.5
| Dimension | SDXL 1.0 | Stable Diffusion 1.5 |
|---|---|---|
| Native image size | 1024 x 1024 | 512 x 512 |
| Architecture | Larger UNet, second text encoder, and additional size and crop conditioning. | Earlier latent diffusion architecture with one text encoder. |
| Inference workflow | Base checkpoint can run standalone or pass latents to an optional refiner. | Single-checkpoint text-to-image workflow in the published v1.5 model card. |
| Migration consideration | Budget for native 1024 workflows and benchmark the optional refiner separately. | Keep when existing 512 pipelines, extensions, or hardware targets are the stronger operational fit. |
| Decision rule | Start here when evaluating SDXL-family image generation. | Retain when the existing SD 1.5 pipeline wins on measured cost, latency, extension compatibility, or output requirements. |
Benchmark the exact workflow you plan to deploy. Compare base-only SDXL, base-plus-refiner SDXL, and the current SD 1.5 pipeline on the same prompts, resolutions, safety controls, and hardware before migrating.
How many parameters does Stable Diffusion XL have?
Hugging Face lists the published SDXL 1.0 base checkpoint at approximately 3 billion parameters. Treat that as the base-checkpoint figure: totals can differ when someone also counts the optional refiner, VAE, and two fixed text encoders.
- The model card describes two fixed text encoders: OpenCLIP ViT-G and CLIP ViT-L.
- The base checkpoint can run by itself; the refiner is a separate checkpoint for final denoising steps.
- State exactly which components are included whenever comparing parameter totals or memory requirements.
Sources: Stability AI SDXL base model card
Run SDXL with Diffusers
The official model card demonstrates loading the base checkpoint with Diffusers, safetensors, and a reduced-precision dtype. It also documents an optional base-plus-refiner pipeline and CPU offload for environments constrained by GPU memory.
- Benchmark base-only and base-plus-refiner output on the same prompts before accepting the extra latency and memory cost.
- Record dtype, scheduler, inference steps, resolution, seed, hardware, and offload settings with every benchmark.
- Use CPU offload when needed, but measure the resulting latency rather than assuming it is free.
Sources: SDXL Diffusers examples
SDXL deployment checklist
SDXL is best treated as a customizable image-generation pipeline rather than a single fixed API. Production readiness depends on the exact checkpoint, adapters, safety controls, watermarking, license review, and evaluation set.
- Evaluate prompt following, text rendering, faces, compositionality, bias, and unsafe-output handling.
- Pin model and dependency versions so a deployment can be reproduced.
- Compare newer hosted and open image models when editing quality, typography, or lower operational overhead matters more than SDXL ecosystem compatibility.
Sources: SDXL limitations and usage
Stable Diffusion XL FAQ
What is Stable Diffusion XL (SDXL)?
Stable Diffusion XL 1.0 is an open-weight latent diffusion model for text-to-image generation at a native 1024 by 1024 resolution. Its base checkpoint can run alone or use an optional refiner during the final denoising steps.
When should teams use SDXL?
SDXL fits self-hosted or customized image-generation workflows where checkpoint control and an open tooling ecosystem matter. Teams should benchmark the exact base or refiner workflow and review license, safety, and infrastructure requirements.
Explore next
Models, tools, and comparisons that connect to this reference.