Damon Who

Posted on Jul 6, 2024 Edited on Sep 13

SDXL + Automatic1111 Setup Guide (2026): Install, VAE & SDXL Turbo

#ai #tutorial #stablediffusion #sdxl

To run SDXL in Automatic1111, put the checkpoint in models/Stable-diffusion and select it in the WebUI. SDXL Turbo uses a separate checkpoint and different sampling settings; an external VAE is optional when the checkpoint's VAE works.

The paths and controls below follow the A1111 documentation and VAE settings. Start with a basic generation, then add the refiner or change the VAE as needed.

Prerequisites for Using SDXL

Start with a working AUTOMATIC1111 installation. The project provides Windows and Linux instructions and a separate Apple Silicon guide. Follow the instructions for your platform before downloading models.

For Windows, the README specifies Python 3.10.6, Git, and webui-user.bat. On Apple Silicon, the documented launcher is ./webui.sh. Use the dependencies in those instructions instead of assuming any Python installation will work.

Hardware settings depend on your GPU. The SDXL optimization wiki includes an 8 GB NVIDIA configuration using --medvram-sdxl --xformers, alongside configurations for smaller cards. That is guidance, not a universal minimum.

Before experimenting, keep a copy of your launch options. Establish whether the ordinary WebUI can load a model and generate an image; this gives you a useful baseline if SDXL introduces a failure.

Updating Your Existing Installation

A1111 added SDXL support in version 1.5.0 and integrated sequential refiner support in 1.6.0. These milestones are documented in the Features wiki; install a build with the capabilities you need.

For a Git installation, close the running WebUI, open a terminal in its installation directory, and run:

git pull

Relaunch with your platform's normal launcher. The Apple Silicon instructions explicitly document this update command and explain that launching the application does not automatically update it.

Keep any local modifications before updating. If Git reports a conflict, resolve that message before continuing; do not treat an interrupted update as a successful installation.

Adding SDXL 1.0 to Automatic1111

Download the single-file checkpoints from Stability AI's repositories:

SDXL 1.0 Base Model on Hugging Face: sd_xl_base_1.0.safetensors.
SDXL 1.0 Refiner on Hugging Face: sd_xl_refiner_1.0.safetensors.

These filenames are also shown in the Diffusers single-file examples. Put the files in stable-diffusion-webui/models/Stable-diffusion, then launch WebUI or refresh its checkpoint list.

Select the base checkpoint first. Stability documents the base as usable by itself; the refiner is an optional model for the final denoising stage. Downloading both does not mean both must run for every image.

Keep checkpoints and standalone VAEs in their respective folders. A VAE download is not a replacement for the base checkpoint, even though both may use the .safetensors extension.

Using SDXL

Basic Text-to-Image Generation:

Open txt2img and choose the SDXL base model from the Stable Diffusion checkpoint menu.
Enter a subject, environment, and visual treatment.
Set the image size to 1024×1024, the standard SDXL size documented by A1111.
Leave the refiner disabled for your initial check and click Generate.

Example prompt: “A hyper-realistic lion in a suit, looking serious, in a Peaky Blinders-inspired environment.”

Optional negative prompt: “sketch, blur, text, logo, monochrome, low contrast, underexposed, overexposed.”

For a starting configuration, use Euler, CFG 5, and 40 steps. This combines Stability's Euler scheduler, the Diffusers guidance default, and the base model card's step example.

Treat that combination as a baseline to evaluate. It is not a claim that every SDXL fine-tune uses the same settings. Save a result you like before changing the sampler, checkpoint, or prompt.

Applying the Refiner in Text-to-Image

In A1111 with integrated refiner support, enable Refiner, choose the refiner checkpoint, and set Switch at. By default, the switch follows model timesteps. The compatibility setting Switch to refiner by sampling steps instead of model timesteps enables the older behavior.

The refiner control defaults to 0.8, matching the split in Stability's base-plus-refiner example. The actual step split depends on the sampler, schedule, and switch mode.

Keep the base selected in the main checkpoint menu. The refiner's checkpoint belongs in the Refiner control. Compare the result with your base-only image before deciding whether the extra stage helps your subject.

Using the Refiner in Image-to-Image

You can also generate with the base, send the image to img2img, and select the refiner for a separate pass. Stability describes this as an alternative to handing partially denoised latents directly between models. Refiner model card

Keep the prompt and dimensions consistent for your first comparison. Start with a modest denoising strength; the SDXL img2img API signature uses 0.3. This is a starting reference, not a guaranteed optimum in A1111.

Lower strength preserves more of the input; higher strength permits greater changes. Review whether the result improves the detail you care about before applying another pass.

Utilizing Predefined Styles

Install StyleSelectorXL:

The extension author's README documents installation through Extensions → Install from URL. Enter the repository URL below, install it, and use the WebUI's restart controls when prompted.

https://github.com/ahgsql/StyleSelectorXL.git

Using StyleSelectorXL:

Find the SDXL Styles panel, enable the extension, write your subject, select a style, and generate. The extension applies positive and negative prompt templates; it does not install a different image checkpoint.

Try styles after you have a usable plain prompt. Keep the unstyled result beside each experiment so you can judge what the template adds. Disable the extension when diagnosing unexpected prompting behavior.

Running SDXL Turbo in Automatic1111

Download sd_xl_turbo_1.0_fp16.safetensors from Stability AI's official Turbo files. The repository also provides sd_xl_turbo_1.0.safetensors. Put the chosen checkpoint in models/Stable-diffusion and select it in txt2img.

Use these initial controls:

Resolution: 512×512.
Sampling steps: 1, with 1–4 as the documented range.
Sampling method: Euler a.
CFG Scale in A1111: 1.
Refiner: disabled; leave the negative prompt empty.

Stability's Turbo model card documents the step range, preferred resolution, and absence of negative-prompt guidance. Its supplied scheduler configuration uses Euler ancestral with trailing timestep spacing.

The guidance numbers differ between interfaces. Stability's Diffusers example sets guidance_scale=0.0. In A1111, CFG 1 yields the conditional prediction without guidance amplification, as follows from its guidance calculation.

That CFG translation follows A1111's implementation. For multiple steps, the SGM Uniform schedule is a practical counterpart to trailing spacing. This schedule mapping is an inference; outputs need not match exactly.

Use Turbo as a standalone workflow without the SDXL refiner. Start by evaluating the direct output before adding other processing stages. Extra stages make it harder to judge whether the Turbo configuration itself is working.

License: read the checkpoint's LICENSE.md before commercial deployment. It contains conditions for commercial use; availability on Hugging Face is not an unrestricted commercial-use grant.

SDXL VAE: which file and where it goes

The VAE converts between pixels and the latent representation used during generation. For txt2img, it decodes the generated latents; img2img also uses it to encode the input image. A1111 VAE explanation

Stability's SDXL VAE provides sdxl_vae.safetensors. Madebyollin's fp16-fix provides sdxl.vae.safetensors, modified to avoid NaNs during half-precision processing. Choose by source and purpose, not filename similarity.

Place a standalone file in stable-diffusion-webui/models/VAE. Open Settings → VAE, choose it in SD VAE, and apply settings. Refresh the list if necessary.

In A1111's settings, None means use the checkpoint's included VAE. Automatic allows automatic selection, including a VAE with a matching checkpoint filename. An explicit selection is clearer when testing a replacement.

Use fp16-fix when the original VAE fails in half precision or when you need a compatible half-precision decoder. A functioning checkpoint VAE does not require replacement just because an external download exists.

Recommended settings

Setting	SDXL Base + Refiner	SDXL Turbo
Initial resolution	1024×1024	512×512
Steps	40 total, following Stability's example	Start at 1; documented range 1–4
Guidance	CFG 5 as a starting reference	A1111 CFG 1; Diffusers guidance 0
Sampler	Euler, matching the base scheduler family	Euler a, matching Turbo's scheduler family
Refiner	Switch at 0.8	Disabled
Negative prompt	Optional	Unused by the documented Turbo workflow
VAE	Included VAE or compatible external SDXL VAE	Begin with checkpoint VAE; diagnose precision if needed

The table assembles the linked model examples and interface translations. Record the configuration alongside each comparison image, especially when moving between base SDXL and Turbo.

Troubleshooting

Black images or a VAE NaN error

Check the terminal message first. If it identifies the VAE, enable Automatically revert VAE to 32-bit floats, select fp16-fix, or test --no-half-vae. A1111 documents the automatic fallback and launch flag.

--no-half-vae keeps the VAE out of half precision. It is more targeted than --no-half, which affects the model. Do not use --disable-nan-check as a repair: that option disables detection rather than correcting invalid values.

Out-of-memory errors

Test --medvram-sdxl for SDXL-specific memory optimization. The wiki describes a tradeoff between memory use and performance. Keep platform-specific choices, such as xformers, aligned with the optimization guide.

Change one setting, restart, and repeat the same generation. A memory error and a VAE precision error are different problems, so keep the error text with your notes.

Check the folder and refresh the relevant list: checkpoints belong in models/Stable-diffusion; VAEs belong in models/VAE. Confirm you downloaded model weights rather than saving the download page.

Advancing with SDXL and Automatic1111

Build a small comparison set around subjects you actually use. Explore prompt engineering, then evaluate one change at a time: style, refiner, or VAE.

FAQ

Does SDXL need a refiner?

No. Stability documents the base as standalone. Add the refiner when your own comparison shows a useful improvement.

Why does SDXL Turbo use CFG 1 in Automatic1111?

A1111's guidance calculation uses CFG 1 for the conditional prediction alone. Stability's Diffusers example disables guidance with 0; the interface values are not interchangeable.

Do I need to download an SDXL VAE separately?

Only if you need an external override. A1111 can use the checkpoint's included VAE; fp16-fix addresses half-precision failures rather than providing a mandatory upgrade.

PromptZone - AI Prompts, Guides and Tools for Builders