SDXL Turbo guide to local generation and sampler settings

To use SDXL Turbo locally, load stabilityai/sdxl-turbo in Diffusers, start at 512×512, disable guidance, and use trailing timesteps with one to four sampling steps. SDXL Turbo is Stability AI's downloadable SDXL model distilled using Adversarial Diffusion Distillation. NVIDIA also documents an A1111 integration through its TensorRT extension. Pipeline settings, model card, A1111 extension

What are the key facts about SDXL Turbo?

What can you test with SDXL Turbo?

Turbo is designed for a small sampling budget. Stability describes generation with one to four steps, which makes prompt and composition experiments a reasonable workflow to evaluate. Model description

For a useful first test, ask for a simple still life with a clear main object. Try a blue ceramic vase on a pale wooden table beside a window. Check whether the object, setting, and lighting all match the request.

Save that result before exploring complicated scenes. A simple reference prompt helps distinguish basic loading problems from a model's difficulty satisfying a more demanding composition.

Diffusers also documents image-to-image use. That provides a separate experiment for testing how a source picture changes under a new prompt; record the source image as well as the text. Image-to-image documentation

If your use case involves previews, judge preview usefulness directly. Ask whether an image helps choose a composition or palette. Do not require a draft to satisfy the same criteria as a finished print asset.

What resolution, guidance, and licensing limits apply?

Turbo's model card identifies problems with legible text, faces, and complete photorealism. Inspect those areas explicitly when deciding whether a result can move beyond a draft. Published limitations

The documented starting resolution is 512×512. Diffusers allows experiments at larger sizes but warns that quality may degrade; larger dimensions should therefore be a separate evaluation, not an assumed improvement. Resolution guidance

Turbo does not use conventional negative prompting in its documented text-to-image recipe. The official example disables guidance, so filling a negative-prompt field should not be treated as a reliable correction method. Usage notes

Describe the wanted scene positively and simplify conflicting instructions. If a portrait keeps including an unwanted prop, remove descriptions that imply it before trying increasingly long exclusion lists.

Its architecture remains SDXL's architecture. Fewer sampling steps do not establish a new small parameter count or guarantee that the model will fit a particular device. Architecture note

For commercial use, check Stability's current licensing page, which includes SDXL Turbo in its Community offering and describes the revenue condition for commercial deployment.

Keep the license decision separate from the inference setup. Licensing FAQ

How do you run SDXL Turbo locally with Diffusers?

1. Use the official stabilityai/sdxl-turbo repository. Read its license and choose the Diffusers loading path before adding optional acceleration. Model card

2. Prepare an environment with CUDA-enabled PyTorch and the documented Diffusers dependencies. The example uses half-precision weights on CUDA; it is not a universal hardware configuration. Pipeline setup

3. Generate a simple baseline image with guidance disabled, the intended schedule, and the documented starting size. Save the output and settings together.
   

import torch
from diffusers import AutoPipelineForText2Image

pipe = AutoPipelineForText2Image.from_pretrained(
    "stabilityai/sdxl-turbo",
    torch_dtype=torch.float16, variant="fp16",
).to("cuda")
pipe.scheduler = type(pipe.scheduler).from_config(
    pipe.scheduler.config, timestep_spacing="trailing"
)
pipe("A blue ceramic vase beside a window", height=512, width=512,
     num_inference_steps=1, guidance_scale=0.0
).images[0].save("turbo-preview.png")

This follows the published guidance to use trailing timesteps and disable classifier-free guidance. Keep the model loaded when comparing generation latency, and report the initial loading time separately. Pipeline recommendations

After the first successful image, compare the supported sampling choices while keeping the brief stable. Note whether additional detail helps the intended deliverable, and retain both accepted and rejected examples.

For image-to-image, the documentation requires the product of step count and strength to be at least one. Its example uses two steps with strength 0.5.

Avoid copying a one-step text-to-image setup without checking this condition. Image-to-image recipe

How do you use SDXL Turbo with A1111 and TensorRT?

NVIDIA's TensorRT extension explicitly lists SDXL Turbo support. It requires installing the extension and generating optimized engines; this is an additional setup path rather than a checkpoint-only procedure. NVIDIA README

In A1111, NVIDIA documents opening Extensions, choosing Install from URL, and entering its repository. Follow the README's engine-generation instructions after installation, including its hardware guidance. Installation instructions

Before attempting that optimization, keep a working baseline and an example image from the ordinary pipeline. Compare the same task after setup, including any preparation time that matters to your usage.

For a node-based alternative, consult the ComfyUI pillar. Save Turbo experiments separately from conventional SDXL graphs so their settings remain identifiable.

How does SDXL Turbo compare with SDXL-Lightning?

The SDXL-Lightning sibling guide explains that model's configuration. Use the SDXL models pillar for the wider checkpoint landscape.

A fair project comparison includes manual correction and final output size. Record those costs instead of turning different inference recipes into a universal speed ranking.

What else should you know about SDXL Turbo?

Does SDXL Turbo use a smaller architecture than SDXL?

SDXL Turbo uses the same architecture as SDXL, according to Hugging Face's pipeline documentation. Its documented sampling recipe uses one to four steps with guidance disabled. Architecture and sampling

Why does my SDXL Turbo negative prompt have no effect?

SDXL Turbo's official recipe does not use negative prompts and sets guidance_scale=0.0. Revise the positive scene description and verify the model's sampling settings before evaluating another result. Model usage

Can SDXL Turbo generate above 512×512?

SDXL Turbo's pipeline accepts larger dimensions, but its documentation warns that quality can degrade. Start at the documented 512×512 size and evaluate larger outputs separately. Resolution guidance

How can I use SDXL Turbo in A1111?

NVIDIA's TensorRT extension documents SDXL Turbo support for A1111. Install the extension and prepare its optimized engines by following NVIDIA's requirements and engine-generation instructions. Official extension

Sources

• Stability AI SDXL Turbo model card
• Hugging Face SDXL Turbo pipeline documentation
• Stability AI generative-models release history
• Stability AI licensing
• NVIDIA TensorRT extension for A1111
• ByteDance SDXL-Lightning model card

Related guides on PromptZone

• Best SDXL Models in 2026
• How to Install and Run SDXL Models in ComfyUI