How to run Stable Diffusion locally in 2026 (the easy way and the classic way)

What you need

No GPU, or a laptop with 4 GB? You can still run SD 1.5 slowly, or generate in the cloud from the same app (see the end of the guide).

The three ways to run Stable Diffusion locally

If you have never done this before, start with A. You can always add B or C later; they can share the same model files.

Route A: the one-click app (no Python)

Imference Desktop is a free, open-source app for Windows and macOS that installs its own isolated inference engine. Nothing touches your system Python.

1. Download the installer for your OS from imference.com/desktop or the GitHub releases page.
2. Run it. The app is not code-signed yet, so you will see a one-time warning. Windows: SmartScreen, click "More info" then "Run anyway". macOS: right-click the app, choose "Open", then "Open" again. You can check the download against the checksums.txt file on the release.
3. Open the app. It installs the engine and detects your GPU on first launch.
4. Pick a model from the catalog. Eight families run locally: SDXL, SD 1.5, Z-Image, FLUX, Chroma, Qwen-Image, Anima and Krea 2. Each one ships pre-tuned (steps, CFG, resolution, quality tags, negative prompt), so the defaults already look right.
5. Type a prompt and hit Generate. The weights download automatically the first time (6 to 13 GB, so the first image takes a few minutes). After that, generation is offline.

What you get beyond the basics: image-to-image, a queue you can stack generations in, model switching mid-run, a gallery that saves every image with its prompt and settings, and "bring your own model": load any .safetensors checkpoint you downloaded from Civitai, pick its family, and it is used in place.

Limits to know: no Linux binary yet (it builds from source), AMD support is a preview, and video generation runs in the cloud, not locally.

Route B: the classic WebUI (AUTOMATIC1111 or Forge)

This is the route most 2023 tutorials describe, and it is still the one with the largest extension ecosystem. The short version:

1. Install Python 3.10.x and tick "Add to PATH".
2. Install Git.
3. Clone the repository:

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

4. Download a model checkpoint (for example SDXL base) into models/Stable-diffusion/.
5. Run webui-user.bat (Windows) or webui.sh (macOS/Linux) and open http://127.0.0.1:7860.

Forge is a drop-in fork with better memory handling; the steps are the same. For a full walkthrough with screenshots, the AUTOMATIC1111 wiki and the install guides on stable-diffusion-art.com are the canonical references. Budget an hour, and expect the first run to download several gigabytes of dependencies on top of the model.

Route C: ComfyUI

ComfyUI is a node editor: you wire loaders, samplers and decoders into a graph. It is the most flexible tool and the reference implementation for new model releases, and there is now a desktop build that removes the Python setup. It is worth learning once you know what you want to build; as a first contact with Stable Diffusion it is overwhelming. We compare the simpler options in 9 ComfyUI alternatives.

Which model should you start with?

• SDXL: the safe default. Good at most subjects, 1024 px, runs on 8 GB.
• SD 1.5: older and less coherent, but the lightest. Use it on 4 to 6 GB cards and for the huge library of community checkpoints.
• FLUX: the best prompt adherence and text rendering, but heavy. Comfortable at 12 GB and up, workable lower with quantized weights.
• Z-Image, Qwen-Image, Chroma, Krea 2: newer families with strong realism or prompt following; try them once SDXL feels limiting.
• Anima: anime and illustration, prompted with tags rather than sentences.

If you are on an Apple Silicon Mac, start with SDXL or SD 1.5; FLUX runs but slowly.

Troubleshooting

• "CUDA out of memory": lower the resolution, generate one image at a time, or switch to SD 1.5. Close games and browsers with hardware acceleration.
• First generation takes forever: it is downloading weights. Watch the progress; the second image will take seconds.
• Windows blocked the installer: SmartScreen on an unsigned app. "More info", then "Run anyway". Verify the checksum if you want to be sure.
• macOS says the app is damaged or can't be opened: right-click, Open. If it still refuses, run xattr -cr "/Applications/Imference Desktop.app" in Terminal.
• AMD card not detected on Windows: ROCm support is a preview; you need Python 3.12 and a recent Adrenalin driver. NVIDIA remains the smooth path.
• Images look wrong or blurry: check that the model's recommended resolution and steps are used. Pre-tuned defaults in the app avoid this; in A1111 set them by hand per model.

Does it work offline?

Yes. Once the app and the weights are on disk, generation needs no internet connection and nothing leaves your machine. That is the whole point for people working with client material or personal photos.

When local is not enough

A 4 GB laptop GPU, a Mac with 8 GB, or a model you don't have room for: the same app has a Cloud toggle that sends the job to imference.com instead, paid per image from $0.003 with a prepaid key or per generation in USDC with no account. Video generation also runs there. You keep one interface and one gallery. If you work inside Claude Code or Cursor rather than an app, the MCP server gives your agent the same catalog.

FAQ