Install Stable Diffusion Locally (A1111 v1.10.1 Guide)

The question every first-timer asks: which Stable Diffusion am I actually installing? There’s the model (weights from Stability AI), the inference code, and the WebUI on top. Most tutorials skip this entirely. They tell you to clone a repo and hope for the best.

This guide installs AUTOMATIC1111’s stable-diffusion-webui v1.10.1 – the go-to desktop interface for open-source image generation – and marks the exact spots where installs break. v1.10.1 was released Feb 9, 2025, and it’s still the current stable tag as of writing. Yes, that’s a long gap between releases. The project is mature; active development lives on dev branches and in forks.

Why bother installing locally

Hosted image generators are faster to start. They’re also rate-limited, paid per image, and they decide what your prompts can contain. Local flips all three. The trade-off: one rough afternoon of dependency wrangling up front.

Local also matters if you plan to fine-tune, use custom checkpoints from Civitai, or batch through hundreds of variations. Cloud APIs charge per generation; that cost compounds fast when you’re iterating on prompts. SD 3.5 itself is free for commercial use under $1M annual revenue via the Stability AI Community License (as of October 2024 announcement) – but only if you’re running it yourself.

System requirements (the honest version)

The README says “NVIDIA recommended” and moves on. Here’s what actually matters:

That Python row isn’t a suggestion. Per the A1111 README: install 3.10.6 – newer versions break the torch integration. Install 3.12, watch the venv build clean, then watch it crash on import. The only fix is uninstalling Python and starting over with 3.10.6. Costs about 20 minutes if you’ve been there before; longer if you haven’t.

On the lighter end, the optimized version of Stable Diffusion can technically run on a GPU with as little as 2.4 GB VRAM – but at that level you’re looking at low resolutions and slow generation times. It works; it’s not fun.

Here’s something worth sitting with before you proceed: the model you choose matters as much as the hardware. SD 1.5 runs on almost anything. SD 3.5 Large (8.1B parameters) wants a serious GPU. SD 3.5 Medium (2.5B parameters) hits the sweet spot for consumer hardware. Getting clear on which model you’re targeting before install saves a round trip through the whole setup.

Installing on Windows (NVIDIA)

Two paths. The zip method is faster; the git clone gives you cleaner updates.

Path A: The zip method (recommended for first install)

1. Go to the v1.0.0-pre release and download sd.webui.zip. This bundle includes Python and git – no global installs needed.
2. Extract somewhere with space. Not Program Files – permissions will cause problems.
3. Double-click update.bat. It pulls the latest WebUI version (v1.10.1 as of this writing).
4. Double-click run.bat. First launch downloads roughly 4 GB of models and dependencies.
5. When you see Running on local URL: http://127.0.0.1:7860, you’re done.

Path B: Manual git clone

# From a non-admin Command Prompt
cd C:AI
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
cd stable-diffusion-webui
webui-user.bat

Edit webui-user.bat before launching if you want xformers (faster generation, lower memory use) – add --xformers to the COMMANDLINE_ARGS= line.

The catch: RTX 50-series GPU? Run switch-branch-tool.bat first to switch to the dev branch. This requirement is documented in the NVIDIA wiki but absent from the main README – which is why 50-series owners hit cryptic errors before finding it.

Installing on Linux

Skip WSL. Per community install guides, WSL adds hours of friction that native Linux avoids entirely.

# Ubuntu / Debian - packages per the official README
sudo apt install wget git python3 python3-venv libgl1 libglib2.0-0

# If your distro ships Python > 3.10, install 3.10 separately
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt install python3.10-venv

# Clone and run
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
cd stable-diffusion-webui
python3.10 -m venv venv
./webui.sh

The official README also lists packages for Fedora (sudo dnf install wget git python3 gperftools-libs libglvnd-glx) and Arch (sudo pacman -S wget git python3).

Getting a model and first configuration

A1111 downloads an older base model on first run – fine for verifying the install, not something you’ll use long-term. Drop any .safetensors checkpoint into models/Stable-diffusion/ and refresh the model dropdown in the top-left of the UI.

For SD 3.5 specifically: grab the 2.5B Medium variant from Hugging Face. It’s the version designed for consumer hardware, released by Stability AI under the Community License (free for commercial use under $1M annual revenue, as of October 2024).

Watch out: If you’re on SD3 or SD3.5, switch your sampler to Euler before generating anything. Turns out v1.10.0 added SD3 support but DDIM and other timestamp samplers don’t work with it – noise, garbled output, no useful error message explaining why. Euler only.

Verifying the install

Open http://127.0.0.1:7860. Type any prompt in the txt2img tab – “red apple on a table” is enough – and hit Generate. An image in 10-60 seconds means the install is clean.

One more check worth doing: look at the console output right at startup. It prints the WebUI version, Python version, and torch version. All three should match what you intended. torch reporting CUDA unavailable on a machine with an NVIDIA GPU means a driver mismatch – not a WebUI bug, and not fixable by reinstalling A1111.

Common errors (from real GitHub Discussions)

• RuntimeError: Couldn't install open_clip – The most-reported first-run crash, caused by outdated pip in the venv. Fix: venvScriptspython.exe -m pip install --upgrade pip, then re-run webui-user.bat.
• Torch is not able to use GPU – Old CUDA drivers, or CPU-only torch got installed. Delete the venv folder and re-run; the launcher rebuilds from scratch.
• Black images on generation – Common on 16-series NVIDIA cards. Add --no-half --precision full to COMMANDLINE_ARGS. Default half-precision math overflows on those cards.
• CUDA out of memory at 512×512 – Close browsers with hardware acceleration (silent VRAM use). Add --medvram or --lowvram for cards under 8 GB.
• Stuck on “Installing requirements” – Not stuck. It’s pulling roughly 5 GB of wheels and the base model. Watch the network monitor.

Upgrading and uninstalling

Git clone? One command: git pull from inside the folder, then relaunch. Zip install? Run update.bat. If an upgrade breaks something, git checkout v1.10.1 puts you back on the known-good tag.

Simple: delete the stable-diffusion-webui folder to uninstall. No registry entries, no system services. The venv lives inside the folder. Back up models/ and outputs/ first – models go with the folder when you delete it.

FAQ

Do I need an NVIDIA GPU, or will AMD / Apple Silicon work?

AMD works – on Linux via ROCm, on Windows via DirectML – but expect more setup friction and slower generation than NVIDIA. Apple Silicon (M1/M2/M3) has its own install path in the wiki and runs well via MPS acceleration. NVIDIA is still the easiest setup, especially if you want to follow community troubleshooting threads without translating every fix.

Why is everyone still on v1.10.1 from February 2025?

The stable tag hasn’t moved. Development happens in forks and extensions now, not core WebUI releases. v1.10.1 covers SD 1.5 through SD3, LoRAs, ControlNet, extensions – there’s no forcing function to upgrade unless a specific feature lands on a new tag.

Is AUTOMATIC1111 still the right choice, or should I look at ComfyUI or Forge?

Honest answer: it depends on your workflow. A1111 is the easiest for prompt-and-generate work. ComfyUI is a node-based graph interface – better for complex multi-pass pipelines and video workflows, steeper initial learning curve. Forge is an A1111 fork tuned for lower-VRAM cards and tends to be faster on the same hardware. Start with A1111. Switch if a specific limitation is actually blocking you – not because someone on a forum said another tool is better.

Next step: grab the SD 3.5 Medium .safetensors from Hugging Face, drop it in models/Stable-diffusion/, set the sampler to Euler, and generate at 1024×1024.