How to Run Stable Diffusion Locally on Windows and Mac

The easiest way to run Stable Diffusion locally in 2026 is Comfy Desktop on Windows and Draw Things on a Mac. Both install like normal apps, with no Python or terminal required. Then you download a model, such as Stable Diffusion 3.5 Medium, and generate images offline.

Here's the short version of how to run Stable Diffusion on your own computer:

• Windows with an NVIDIA or AMD card: install Comfy Desktop, the official ComfyUI app.
• Apple silicon Mac: use Draw Things (free on the App Store) or Comfy Desktop.
• Hardware: Stability AI says SD 3.5 Medium needs about 9.9 GB of VRAM for full performance. ComfyUI can stream weights to get by on much less, just more slowly.
• Avoid starting on AUTOMATIC1111 or DiffusionBee unless you already know them. Neither has tagged a new version since mid-2024.

Below: what each front-end looks like as of October 2026, the hardware you need, and the setup steps for each platform.

Which Stable Diffusion app to use in 2026

Most tutorials still point to tools from 2023. Several of those have gone quiet. We checked each project's own GitHub repo or store listing on October 30, 2026:

ComfyUI is now where most of the work happens. It's a node-based interface: each step (load model, write prompt, sample, save) is a box you wire together. That looks intimidating at first. In practice you open a ready-made workflow, type a prompt and press run. The project's README says it supports NVIDIA, AMD, Intel, Apple silicon and Ascend hardware through a manual install.

AUTOMATIC1111's Stable Diffusion WebUI was the classic tab-based interface. It still works, and a small dev-branch fix landed in March 2026, but there has been no new version since v1.10.1, tagged in 2024. Its successor, Forge, says in its own README that the original WebUI "is almost static now." Forge itself hasn't had a commit since mid-2025.

Two other names come up often. InvokeAI is still actively developed and installs through a launcher app. Fooocus describes itself as being in "limited long-term support" with bug fixes only, and it's built entirely on SDXL. Its README points people who want newer models like FLUX to Forge or ComfyUI.

One more change trips up older guides. The original ComfyUI desktop repo is archived and marked "Superseded by" the new Comfy Desktop project. If a tutorial links the old one, use the new app instead.

Hardware requirements for running Stable Diffusion locally

The honest answer depends on the model and the app. Here's what each official source actually says:

Two practical takeaways:

1. VRAM decides speed, not whether it runs. With 6–8 GB of VRAM you can still generate. ComfyUI moves parts of the model between your graphics card and system memory. It works, it's just slower.
2. Disk fills up fast. The app itself is a few gigabytes. Each model checkpoint adds several more. Plan for tens of gigabytes if you start collecting models.

On a Mac, there's no separate VRAM. Apple silicon shares one pool of memory between the CPU and GPU, so a 16 GB Mac is a much more comfortable starting point than an 8 GB one.

How to run Stable Diffusion on Windows

The easy route: Comfy Desktop

1. Download the installer from comfy.org/download and run it like any Windows app.
2. Pick Local when the first-run screen asks how you want to run Comfy. (Cloud runs on Comfy's servers instead, which defeats the point of a local setup.)
3. Tick Express Install to accept the recommended settings. The app sets up its own Python and GPU libraries, so you don't install either yourself.
4. Download a model. Start with a Stable Diffusion checkpoint from Hugging Face (see the model section below). Comfy Desktop keeps models in a shared library at %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared, so every install you create can use them.
5. Open a text-to-image workflow, choose your model in the "Load Checkpoint" box, type a prompt and run it.

The Comfy Desktop docs walk through each screen if anything looks different on your machine.

The portable route for NVIDIA, AMD and Intel cards

If you'd rather not install anything, ComfyUI also ships a portable Windows build. You download a 7z archive, extract it and run it. The README lists separate builds for NVIDIA, AMD and Intel GPUs. Models go in ComfyUI\models\checkpoints.

Two details from the README that save time:

• The standard NVIDIA build supports 20-series cards and newer. Older 10-series cards need the separate CUDA 12.6 build.
• For a manual install on an AMD card, the README's Windows ROCm instructions ask for Windows 11, a current AMD driver and 64-bit Python 3.13.

The README itself says the portable build "is not recommended for regular users" and points people to the desktop app. Use it if you know why you want it.

If you already use AUTOMATIC1111 or Forge

They still run. The A1111 README asks for Python 3.10.6 specifically, and says newer Python versions don't support its torch version. That single requirement causes most failed installs. ComfyUI's manual install, by contrast, says Python 3.13 is very well supported. If you're starting fresh, skip it and use ComfyUI.

How to run Stable Diffusion on a Mac

Draw Things: the simplest option

Draw Things is a native app for Mac, iPhone and iPad. Its site says it's "completely free and runs offline." The App Store lists it as free with in-app purchases, and optional cloud compute is available if you want it. The App Store listing requires macOS 12.4 or later.

How to use it:

1. Install Draw Things from the Mac App Store.
2. Open the model menu and download a model from inside the app.
3. Type a prompt and generate. Everything after the model download runs locally.

Stability AI's own Hugging Face page for SD 3.5 Medium lists Draw Things and DiffusionBee under its local apps.

Comfy Desktop on Apple silicon

If you want the same workflows Windows users share, install Comfy Desktop on your Mac too. The docs list macOS 13 or later on Apple silicon. Setup is identical: download, choose Local, Express Install, add a model.

For a manual install instead, the ComfyUI README says to install the PyTorch nightly first, following Apple's Accelerated PyTorch on Mac guide, then follow the normal manual steps.

DiffusionBee and Intel Macs

DiffusionBee was the original one-click Mac app, and it still works. But its last GitHub release dates to August 2024, so don't expect quick support for the newest models. Its site says Intel Macs are supported but "very slow compared to M1 machines" without a dedicated graphics chip.

If you're on an Intel Mac, be realistic. Comfy Desktop's Mac build is Apple silicon only, and image generation on an Intel CPU is slow. A hosted tool is usually the better fit. Our guide to AI on Mac covers which layers make sense on which machine.

Which Stable Diffusion model to download

The app is only half of the setup. You also need model weights, and the licence matters if you plan to use images commercially.

Stable Diffusion 3.5 Medium is the best starting point for most machines. Stability AI's SD 3.5 announcement describes it as a 2.5-billion-parameter model designed to run "out of the box" on consumer hardware. Note two things:

• It's gated. You sign in to Hugging Face and agree to share your username and email with Stability AI before you can download it.
• It uses separate text encoders. That's why Stability quotes its 9.9 GB figure "excluding text encoders." The model card recommends ComfyUI for local use and links the right files.

Stable Diffusion 3.5 Large has 8.1 billion parameters. Stability positions it for professional use at 1 megapixel, so expect to need a high-end card. SD 3.5 Large Turbo is a distilled version that generates in four steps.

SDXL 1.0 is older but has the largest library of community fine-tunes and LoRAs. Its model card uses the CreativeML Open RAIL++-M licence.

Can you use the images commercially?

For SD 3.5, the answer is in the Stability AI Community License. It keeps the models free for people and organizations with annual revenue under US $1 million, including commercial use (which requires registration). Above that threshold, you need a separate licence from Stability AI. The agreement also says you own the outputs, "to the extent permitted by applicable law."

Want a comparison with other open models, like FLUX? Our roundup of Midjourney alternatives compares their licences side by side.

Common problems and fixes

Running image models vs running AI agents locally

Stable Diffusion is a single model that takes a prompt and returns an image. Running an AI agent locally is a different job: you need a language model, a runtime, an agent loop and tools. We cover that in how to run AI agents locally.

And if you only want free images without managing a GPU, the hosted options in our best free AI image generators list are faster to start with.

Either way, the hard part is the same: getting someone else's working setup to run on your machine. Taku is a desktop workspace where you can run AI apps and workflows that others have already set up, then save the ones you keep using. Taku is in Beta, and the Mac app is available now.

FAQ

How do I run Stable Diffusion locally for free?

Install Comfy Desktop (Windows or Apple silicon Mac) or Draw Things (Mac), then download a model from Hugging Face. The apps are free, and SD 3.5 is free to use under the Stability AI Community License if your annual revenue is under US $1 million.

How much VRAM do I need to run Stable Diffusion?

Stability AI says SD 3.5 Medium needs 9.9 GB of VRAM, excluding text encoders, for full performance. ComfyUI says it can run large models on as little as 4 GB of VRAM plus 8 GB of RAM by streaming weights, at lower speed.

Can I run Stable Diffusion on a Mac?

Yes. On Apple silicon, Draw Things and Comfy Desktop both run it locally. Intel Macs can run DiffusionBee, but its own site warns it is very slow without a dedicated graphics chip.

Is AUTOMATIC1111 still maintained?

Barely. The latest version is v1.10.1, tagged in July 2024, and the default branch hasn't changed since. A dev-branch fix landed in March 2026. Forge, its main fork, has had no commits since June 2025.

Can I run Stable Diffusion without a GPU?

ComfyUI's docs say a dedicated GPU is recommended but not required, and its portable Windows build can run on the CPU. Expect each image to take much longer.

Is Stable Diffusion safe to run locally?

The model runs offline on your machine, so prompts and images stay local. The main risk is custom nodes and extensions, which are third-party code. Install only ones you trust.