Stable Diffusion XL.
Stable Diffusion XL is an open-weight image model from Stability AI. Windpaint doesn't serve it yet, so there's no model id or price. Request early access to start with the live models.
- Status
- Coming soon
- Task
- Text to image
- Input
- Prompt, optional negative prompt
- Output
- Images
- Developer
- Stability AI
- Parameters
- 2.6B UNet
- Licence
- CreativeML Open RAIL++-M
- Released
- July 2023
What the model does.
Paraphrased from the official model card.
Stable Diffusion XL 1.0 is a latent diffusion text-to-image model from Stability AI that uses two CLIP text encoders. The base model can run on its own, or hand its latents to a separate refiner model for the last denoising steps.
It's older than the other models here, but it has a large set of community fine-tunes, LoRAs and ControlNets, and many existing pipelines are built on it.
Coming soon to Windpaint.
Stable Diffusion XL isn't served, so it has no model id, no price and no request example here. When it goes live, this page gets its tiers, prices and request fields like the live models have.
Want this model? Request early access. We're adding models as we add capacity.
Live for images today: Z-Image Turbo (text to image) and FLUX.2 [klein] 4B (text to image). See Image generation for what the API does now.
Is Stable Diffusion XL licensed for commercial use?
Licence: CreativeML Open RAIL++-M.
The Open RAIL++-M licence has no non-commercial restriction and permits hosting the model as a service. The use-based restrictions in its Attachment A apply, and a host has to put them into the terms it agrees with its users.
- The licence counts "providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access" as Distribution. A host must include the paragraph 5 use-based restrictions (Attachment A) as an enforceable provision in the agreement governing users' use, tell users the model is subject to paragraph 5, and require users to comply with it. stable-diffusion-xl-base-1.0, LICENSE.md
- Anyone who receives the model gets a copy of the licence, modified files must carry prominent change notices, and copyright, patent, trademark and attribution notices must be kept. Derivatives must carry at least the same use-based restrictions. stable-diffusion-xl-base-1.0, LICENSE.md
- Outputs: the licensor claims no rights in the output you generate. You are accountable for outputs and their later use, and no use of output may contravene the licence. stable-diffusion-xl-base-1.0, LICENSE.md
- Attachment A bans, among other things: illegal use, harming minors, verifiably false information meant to harm others, harmful PII, defamation or harassment, fully automated decisions affecting legal rights, discrimination, exploiting vulnerable groups, medical advice, and uses in justice, law-enforcement, immigration or asylum processes. stable-diffusion-xl-base-1.0, LICENSE.md
- The licensor reserves the right to restrict usage of the model that violates the licence, "remotely or otherwise". stable-diffusion-xl-base-1.0, LICENSE.md
- The card's "Direct Use" section says the model "is intended for research purposes only". The licence itself has no research-only or non-commercial clause and explicitly permits hosting for third-party access, but it also says it "is informed by the model card". stable-diffusion-xl-base-1.0 model card
This is our reading of the licence, not legal advice. Read the licence text before you rely on it.
Stable Diffusion XL hardware requirements.
Published by Stability AI for running the weights yourself. These are the developer's figures, not Windpaint's.
- Published VRAM
- Not stated
- Applies to
- No VRAM figure on the card. fp16 weights; CPU offload if VRAM-limited.
- The card gives no VRAM figure. It loads the weights in float16 (
variant="fp16"). If you're limited by GPU VRAM, the card says to callpipe.instead ofenable_model_cpu_offload() .to("cuda"). stable-diffusion-xl-base-1.0 model card - The card also documents running it through Optimum with OpenVINO and ONNX Runtime pipelines. stable-diffusion-xl-base-1.0 model card
Stable Diffusion XL native resolution.
What the developer's own code and examples use.
- Stability's SDXL report says the final stage used multi-aspect training at about 1024 × 1024 total pixel area, with height and width varied in multiples of 64. Its bucket list includes 1024 × 1024, 1152 × 896 and 896 × 1152. arXiv 2307.01952
Recommended settings for Stable Diffusion XL.
Steps, guidance and defaults from the developer's card and code, for running the weights yourself.
- Base plus refiner (ensemble of experts) example: 40 total steps. The base handles the first 80% (
denoising_end=0.8,output_type="latent") and the refiner the rest (denoising_start=0.8), on the same prompt. stable-diffusion-xl-base-1.0 model card - The card says the base model can be used as a standalone module. stable-diffusion-xl-base-1.0 model card
- Requires diffusers >= 0.19.0, plus invisible_watermark, transformers, accelerate and safetensors. On torch >= 2.0, the card suggests wrapping the UNet with
torch.compile. stable-diffusion-xl-base-1.0 model card
Stable Diffusion XL limitations.
As the developer states them.
- Does not achieve perfect photorealism. stable-diffusion-xl-base-1.0 model card
- Cannot render legible text. stable-diffusion-xl-base-1.0 model card
- Struggles with compositional prompts such as "A red cube on top of a blue sphere". stable-diffusion-xl-base-1.0 model card
- Faces and people in general may not be generated properly, and the autoencoding part is lossy. stable-diffusion-xl-base-1.0 model card
- Stability's report adds trouble with intricate structures such as human hands, and "concept bleeding", where distinct visual elements merge or overlap. arXiv 2307.01952
Questions.
Can I call Stable Diffusion XL on Windpaint today?
No. Stable Diffusion XL isn't served yet, so it has no model id, no price and no entry in the catalog endpoint. For images today, z-image-turbo and flux-2-klein-4b are live. There's no date for Stable Diffusion XL.
What resolutions does SDXL support?
Stability's report says SDXL was fine-tuned on several aspect ratios at about 1024x1024 total pixels, such as 1024x1024, 1152x896 and 896x1152.
Can I use SDXL commercially?
The CreativeML Open RAIL++-M licence has no non-commercial clause and explicitly allows hosting the model as a service. Its Attachment A use restrictions apply, and anyone hosting it has to put those restrictions into its terms with users.
What licence is Stable Diffusion XL under?
CreativeML Open RAIL++-M.
Can SDXL render text?
The model card lists "cannot render legible text" as a limitation.
Is Stable Diffusion XL open source?
No, it's open-weight. The weights are downloadable, but its licence (CreativeML Open RAIL++-M) restricts how they can be used, so it isn't an open-source licence.
How do I ask for Stable Diffusion XL?
Request early access at app.windpaint.ai/register. We're letting people in gradually while we add models and capacity.