AUTOMATIC1111: install Stable Diffusion WebUI (2026)

#What is AUTOMATIC1111?

Stability AI announced the public release of Stable Diffusion on August 22, 2022. A developer known by the pseudonym AUTOMATIC1111 created a repository on GitHub that same day: a browser interface built with the Gradio library and released under the AGPL-3.0 license, eliminating the need to type Python commands. The repository now has more than 160,000 stars. The interface offers tabs for text-to-image, image-to-image, inpainting, and upscaling, plus a huge extension library.

Like all local tools, it runs on your machine and opens in the browser at http://127.0.0.1:7860, never leaving your local network by default. Prompts, settings, and images remain on your disk, available for offline access and reuse.

• Stable Diffusion: versions and which one to choose

#What you need

AMD cards work natively on Linux with ROCm. On Windows, the official wiki is clear: “Windows+AMD support has not officially been made for webui,” and the practical route goes through lshqqytiger’s community DirectML fork, where training was still not working according to the wiki page, edited in January 2024. On Mac Apple Silicon, the official wiki documents installation via Homebrew and the webui.sh script: most features work, but training is very slow and memory-intensive. NVIDIA remains the best-supported path for the project.

#Installation on Windows and Linux

1. 01
   Install Python 3.10.6

   From python.org, by checking “Add Python to PATH.” The README states that newer versions do not support torch: stick with 3.10.6, even though the same README allows 3.11 for some very recent Linux distributions.
2. 02
   Install Git

   Git for Windows on Windows; your distribution's git package on Linux.
3. 03
   Clone the repository and run the script

   In a folder with a short path, without spaces or accented characters. On Windows, the script is webui-user.bat; on Linux, it is webui.sh.
4. 04
   Wait during the first launch

   The script downloads PyTorch and its dependencies, several gigabytes in total: allow ten minutes to half an hour. When a local address appears, open it in your browser.

git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
cd stable-diffusion-webui

# Windows
webui-user.bat

# Linux (après : sudo apt install git python3.10-venv libgl1 libglib2.0-0)
./webui.sh

#Where to store the models

models/Stable-diffusion

Checkpoints—in other words, the main model files—in .safetensors format.

models/Lora

LoRAs, which add a style or subject to a base model.

models/VAE

Optional VAEs; many SDXL checkpoints include their own.

extensions

The extensions, managed from the tab with the same name.

outputs

Your images, sorted by mode and date.

A fresh installation may not contain a model: older versions automatically downloaded SD 1.5 on first launch, but that is no longer guaranteed. Download an SDXL checkpoint from Hugging Face, place it in models/Stable-diffusion, then click the refresh icon next to the model selector. Use only .safetensors files: the older .ckpt format can execute code when opened, as detailed below.

• Hugging Face: find and verify a model file

#Launch options based on your VRAM

The options go on the COMMANDLINE_ARGS line in webui-user.bat on Windows or webui-user.sh on Linux.

set COMMANDLINE_ARGS=--xformers --medvram-sdxl --autolaunch

#Settings for a first image

Describe rather than give commands, in English, with the subject first: “portrait of an elderly fisherman, window light, 85mm, shallow depth of field” works better than a list of isolated keywords. The negative prompt lists what to avoid, for example “blurry, deformed hands, extra fingers, low quality” to filter out the most common defects. One function to learn early: the PNG Info tab rereads the prompt and all settings from any image produced by the interface, allowing you to reproduce or adjust a successful generation without retyping everything.

#A1111, Forge, or ComfyUI in 2026

The verdict is simple. Choose A1111 if you’re following a course, tutorial, or extension written for it, or if your current setup does what you need with SD 1.5 and SDXL. For a new installation with this interface style, look at Forge: it’s a fork maintained by lllyasviel, the author of ControlNet and Fooocus, designed for better resource management and supporting Flux, while preserving the same workflow. The Forge repository says it itself: because the original project is “almost static now,” Forge resynchronizes with it “every 90 days, or when important fixes.” But the Forge repository itself has not received a commit on its main branch since June 2025. Development is now carried on by a community fork, Forge Neo by Haoming02, whose README announces support for recent models (Qwen-Image, Wan 2.2, Z-Image) and which was still being updated in late September 2026. For a community project, check its activity before committing to it. Choose ComfyUI for current models, video, or reproducible pipelines.

• ComfyUI: installation and getting started
• Fooocus: the simplest option for getting started
• The AI Image Kit: settings by graphics card

#Useful extensions and model security

#ADetailer: automatically fix faces and hands

In a wide-shot generated image, faces and hands remain Stable Diffusion's weak point. The ADetailer extension, whose official repository describes it as an extension that “does automatic masking and inpainting,” automatically detects these areas with a recognition model and then runs targeted inpainting on them with its own prompt. It installs directly from the Extensions tab by pasting its GitHub address into “Install from URL.”

Inpainting, specifically, is one of the interface's four basic modes, along with text-to-image, image-to-image, and upscaling: it regenerates only a masked area of the image while keeping the rest identical. ADetailer automates this step instead of making you paint the mask by hand on every face.

#.ckpt and .safetensors files: what the format changes

A .ckpt file is a Python serialization in pickle format, which can execute arbitrary code when opened. Malicious checkpoints have already circulated on community platforms. The .safetensors format solves the problem at its root: according to the format’s official repository, it “implements a new simple format for storing tensors safely (as opposed to pickle) and that is still fast (zero-copy)”. AUTOMATIC1111 includes a check that blocks loading a pickle deemed suspicious before deserializing it, but the simplest rule remains to download only .safetensors files, especially from sources other than Hugging Face or Civitai.

• ControlNet: mastering image composition
• Source: ADetailer’s official repository
• Source: the .safetensors format explained by Hugging Face

#Common errors

CUDA out of memory

The image size or model exceeds the VRAM. Add --medvram or --medvram-sdxl, lower the resolution, and close other applications that use the GPU.

Torch is not able to use GPU

Driver too old, or GPU that is not a NVIDIA. Update the driver; the CUDA Toolkit is not required.

Installation fails on dependencies

Wrong Python version. Install 3.10.x, then delete the venv directory before trying again.

A recent RTX 50 is not being used

The stable version bundles a PyTorch release from before the Blackwell architecture. A project contributor documents the workaround in a GitHub discussion updated May 1, 2025: switch to the dev branch (git switch dev), which uses PyTorch 2.7.0 with RTX 50 support, then run it once with --reinstall-torch in COMMANDLINE_ARGS and remove the option afterward. Otherwise, ComfyUI.

• CUDA: driver, Toolkit, versions, and what you really need to install
• The official GitHub repository

#FAQ