AUTOMATIC1111 Stable Diffusion WebUI: Install Guide, Key Settings, and Whether to Still Use It

What AUTOMATIC1111 is

When Stable Diffusion's weights were released in August 2022, running them meant typing Python commands. Within weeks a developer using the handle AUTOMATIC1111 published a browser-based interface built on the Gradio library, under the AGPL-3.0 license. It grew into the most widely used image-generation front end of the following two years. The project lives at github.com/AUTOMATIC1111/stable-diffusion-webui.

Like every local tool, it runs on your machine and opens in your browser at http://127.0.0.1:7860. Prompts and images stay on your disk.

What you need

	Minimum	Comfortable
GPU	NVIDIA with 4 GB of VRAM (SD 1.5 only, with memory-saving flags)	NVIDIA with 8 GB for SDXL; 12 GB+ for SDXL with LoRAs, ControlNet and upscaling
System RAM	8 GB	16 GB or more
Disk	≈ 12 GB for the install plus one model	50 GB+ on an SSD; model collections grow quickly
Software	Python 3.10.x and Git	The exact Python minor version matters: newer versions break dependencies
Other hardware	AMD cards work on Linux (ROCm) and, with more effort, on Windows; Apple Silicon works but is slow. NVIDIA is the supported path.

To check your card's memory, see what VRAM is and how to check it.

Installation

Windows (NVIDIA)

1. Install Python 3.10.6 from python.org and tick "Add Python to PATH." Do not use 3.11 or later.
2. Install Git for Windows.
3. Open a terminal in a folder with a short path and no spaces, and run:

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

4. The first launch downloads PyTorch and dependencies, several gigabytes, and can take 10 to 30 minutes. When it prints a local URL, open it in your browser.

Linux

sudo apt install git python3.10-venv libgl1 libglib2.0-0
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
cd stable-diffusion-webui
./webui.sh

On distributions that ship a newer Python, install 3.10 alongside it and point the script to it with python_cmd="python3.10" in webui-user.sh.

Where files go

Folder	Contents
models/Stable-diffusion	Checkpoints: the main model files (.safetensors)
models/Lora	LoRAs, which add a style or subject to a base model
models/VAE	Optional VAE files; many SDXL checkpoints include their own
embeddings	Textual inversion embeddings
extensions	Installed extensions (managed from the Extensions tab)
outputs	Your images, sorted by mode and date

A fresh install has no model; older versions downloaded SD 1.5 automatically, but do not count on it. Download an SDXL checkpoint from Hugging Face, place it in models/Stable-diffusion, and press the refresh icon next to the checkpoint selector. Our Hugging Face guide explains how to find and vet model files. Use .safetensors only: the older .ckpt format can execute code when loaded.

Launch flags by VRAM

Options go in the COMMANDLINE_ARGS line of webui-user.bat (Windows) or webui-user.sh (Linux).

Flag	Effect	Use it when
--xformers	Memory-efficient attention: faster, less VRAM	Almost always on NVIDIA cards
--medvram	Keeps only part of the model on the GPU at a time; slower	6–8 GB cards running SDXL
--medvram-sdxl	Same, applied to SDXL models only	8 GB cards that handle SD 1.5 fine
--lowvram	Aggressive offloading; much slower	4 GB cards, as a last resort
--api	Enables the REST API	Driving the WebUI from scripts or other apps
--listen	Accepts connections from other devices on your network	Using it from a tablet or another PC; add authentication
--autolaunch	Opens the browser automatically	Convenience

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

The settings that matter for a first image

Setting	Starting value	What it does
Sampling method	DPM++ 2M Karras, or Euler a	The denoising algorithm; these two are reliable defaults
Sampling steps	20–30	More steps, more time, diminishing returns past 30
Width × height	1024 × 1024 for SDXL, 512 × 512 for SD 1.5	Stay near the model's native resolution or compositions fall apart
CFG scale	5–7	How strictly the image follows the prompt; too high looks burnt
Seed	-1 (random)	Fix it to reproduce an image while changing one thing
Hires. fix	Off at first	Generates small then upscales with a second pass; the classic way to get detail

Describe rather than command, in English, subject first: "portrait of an elderly fisherman, window light, 85mm, shallow depth of field." The negative prompt lists what to avoid. One feature worth knowing early: the PNG Info tab reads the prompt and every setting back from any image the WebUI generated.

A1111 vs Forge vs ComfyUI in 2026

	AUTOMATIC1111	Forge	ComfyUI
Interface	Tabs and forms	Same tabs and forms	Node graph
Learning curve	Gentle	Gentle	Steep
Speed and VRAM efficiency	Baseline	Noticeably better, especially under 12 GB	Best
Newer models (Flux and later)	No	Flux supported	Nearly everything, usually first
Video models	No	No	Yes
Extensions and tutorials	The largest back catalog	Most A1111 extensions work	Large, different ecosystem (custom nodes)
Development pace	Slow: last major release line dates from mid-2024	Irregular but more recent	Very active

The verdict is straightforward. Choose A1111 if you are following a course, a tutorial or an extension written for it, or if an existing install does what you need with SD 1.5 and SDXL. Choose Forge for a new install if you like this style of interface: it is a fork by the author of ControlNet and Fooocus with better memory management, and your habits transfer unchanged. Choose ComfyUI if you want current models, video, or reproducible pipelines; start with our ComfyUI beginner's guide. For settings per GPU and ready-made workflows, see the Local Image AI Kit.

Common errors

Message or symptom	Cause	Fix
CUDA out of memory	Image size or model exceeds VRAM	Add --medvram or --medvram-sdxl; lower the resolution; close other GPU apps
Torch is not able to use GPU	Driver too old, or a non-NVIDIA GPU	Update the NVIDIA driver; you do not need the CUDA Toolkit (see what is CUDA)
Install fails on dependencies	Wrong Python version	Use Python 3.10.x and delete the venv folder before relaunching
Black or green images	Half-precision problems on some cards or VAEs	Add --no-half-vae
A recent RTX 50 card is not used	Bundled PyTorch predates the architecture	Upgrade PyTorch inside the venv to a build that supports your GPU, or use Forge or ComfyUI

Hardware data referenced across BestLLMfor is open through our public API (CC BY 4.0) and MCP server. The WebUI's own wiki documents every flag, and the Forge fork lives at github.com/lllyasviel/stable-diffusion-webui-forge.