How to Run Stable Diffusion Locally in 2026
A step-by-step guide to running Stable Diffusion on your own Windows, Mac or Linux computer, from hardware and installers to models and your first image.
Key takeaways
The easiest route in 2026 is Comfy Desktop, a free app for Windows and macOS that installs ComfyUI with Python and PyTorch already set up for your graphics card.
SD 1.5 runs on a graphics card with 4GB of video memory and SDXL wants 8GB. Apple Silicon Macs work too, ideally with 16GB of memory or more.
Download models as .safetensors files. Skip .ckpt files, which can run malicious code when they're loaded.
SD 3.5 is still Stability AI's newest open image model. It's free for personal use and for commercial use under $1 million in annual revenue, while SD 1.5 and SDXL use OpenRAIL licenses that allow commercial use with some use restrictions.
Automatic1111, once the default interface, hasn't been updated since July 2024. ComfyUI, Forge Neo and Stability Matrix are the ones to use now.
What hardware do you need for Stable Diffusion?
A recent NVIDIA graphics card is the smoothest path, but AMD cards and Apple Silicon Macs work too. Here's what each Stable Diffusion model needs at a minimum.
Minimum requirements by model
Model | Minimum video memory | Minimum RAM | Download size |
|---|---|---|---|
SD 1.5 | 4GB | 8GB | 4.3GB |
SDXL 1.0 | 8GB | 16GB | 6.9GB |
SD 3.5 Medium | 9.9GB, not counting text encoders | Not published | 5.1GB plus text encoders |
SD 3.5 Large | Not published | Not published | 16.5GB plus text encoders |
Notes for NVIDIA, AMD and Mac
The SD 1.5 and SDXL minimums come from InvokeAI's system requirements, and the SD 3.5 Medium figure is Stability AI's own. Cards with less video memory often still work, though they're slower. ComfyUI says its weight streaming lets even big models run on 4GB cards with 8GB of RAM.
On NVIDIA, the 20 series and newer get the current builds. ComfyUI keeps a separate portable build for 10 series cards.
On AMD, ComfyUI now runs natively on Windows 11 through ROCm for RDNA 2 cards and newer, which means the Radeon RX 6000 series onward. On Linux, you'll install the ROCm version of PyTorch.
On a Mac, any Apple Silicon chip works, but Stable Diffusion slows down badly when it's short of memory, so InvokeAI recommends 16GB or more.
Step 1: Pick an interface
Stable Diffusion is a model, so you need an app to run it. These are the options worth installing in 2026.
ComfyUI: builds images from a graph of connected nodes, supports most new open image models, and follows a weekly release cycle. It's the best choice for most people.
Forge Neo: keeps the classic tabbed WebUI layout with sliders, and it's actively maintained, with support for newer models such as Krea 2.
Stability Matrix: a free manager for Windows, Linux and Apple Silicon Macs that installs ComfyUI, Forge Neo, InvokeAI, SwarmUI and others in one click.
Draw Things: a free app for Mac, iPhone and iPad that runs offline with almost no setup.
InvokeAI: a canvas-first app for inpainting and outpainting, with its own launcher.
Step 2: Install ComfyUI
The steps differ a little by platform.
On Windows or a Mac
Download Comfy Desktop from comfy.org/download. It needs Windows 10 or later, or macOS 13 Ventura or later on an Apple Silicon Mac, and each install takes about 4.85GB of disk space before you add any models.
The app bundles its own Python with PyTorch and the GPU libraries prebuilt, so you don't have to install Python or Git yourself. It can also keep several ComfyUI installs side by side and roll one back to an earlier snapshot if an update breaks something.
On Linux, or for a manual install
Install the comfy-cli tool with pip, then run comfy install. ComfyUI supports Python 3.13 well and suggests 3.12 if a custom node gives you trouble.
NVIDIA cards from the 20 series up need a PyTorch build for CUDA 13.0 or newer. AMD cards on Linux use the ROCm build of PyTorch instead.
Step 3: Download a model
SDXL 1.0 is the best first model. It's a free 6.9GB download with no sign-up and runs on 8GB cards. It also has a large library of community fine-tunes.
SD 3.5 Large and Medium ask you to log in to Hugging Face and agree to share your contact information before you can download them, and their main model files don't include the text encoders, which you download separately from the same page.
Put checkpoint files in ComfyUI's models/checkpoints folder. Forge Neo uses models/Stable-diffusion instead. Always pick the .safetensors version of a model. The older .ckpt format is a Python pickle file, and Hugging Face warns that loading a pickle can run arbitrary code on your computer.
Most community fine-tunes live on Civitai. Since April 2026 it's run two sites, civitai.com for safe-for-work models and civitai.red for everything else, and one account works on both.
Stable Diffusion licenses
Model | License | Commercial use |
|---|---|---|
SD 1.5 | CreativeML OpenRAIL-M | Allowed, with use restrictions |
SDXL 1.0 | CreativeML OpenRAIL++-M | Allowed, with use restrictions |
SD 3.5 Large, Medium and Large Turbo | Stability AI Community License | Free under $1 million in annual revenue, with registration for commercial use |
Step 4: Make your first image
ComfyUI's default workflow is a basic text-to-image graph. Pick your model in the Load Checkpoint node, type what you want in the CLIP Text Encode node wired to the sampler's positive input, and click Run.
Generate at the size each model was trained on: 512 by 512 pixels for SD 1.5 and 1024 by 1024 for SDXL. Other sizes work, but they'll tend to look worse the further you stray.
How to fix common first-run problems
Four problems come up a lot on a first run.
Out of memory: lower the resolution first. In ComfyUI, the novram flag is the fallback when a model still won't fit, and in Automatic1111 the medvram and lowvram flags trade speed for memory.
Black or green images: some cards can't handle half-precision math in the VAE, the part that turns the result into pixels. Use the no-half-vae flag in Automatic1111 or fp32-vae in ComfyUI. For SDXL, the community's fp16-fix VAE solves it too.
Slow generation on a Mac: check the memory pressure graph in Activity Monitor. If it's red, close other apps or drop to a smaller model.
Speed add-ons: you don't need them. Forge Neo's README says PyTorch's built-in attention is usually as fast as xformers and more stable.
Stay safe with custom nodes
ComfyUI's custom nodes are Python code that runs with full access to your computer, and some have been malicious. In 2024 a node called ComfyUI_LLMVISION stole browser passwords and credit card details. In October 2025, two upscaler packs carrying the Akira Stealer malware were downloaded 790 times before Comfy banned them from its registry.
Install nodes from well-known authors only, and don't open ComfyUI to the internet with the listen flag unless you've put a password in front of it. In April 2026, Censys reported an active campaign hijacking exposed ComfyUI servers for cryptomining, with more than 1,000 instances open to the internet.
Frequently asked questions
Can I run Stable Diffusion without a graphics card?
Yes, but slowly. ComfyUI has a CPU mode, and Apple Silicon Macs use their built-in GPU. A dedicated graphics card is far faster.
How much VRAM do I need for Stable Diffusion?
4GB for SD 1.5 and 8GB for SDXL, per InvokeAI's requirements. Stability AI says SD 3.5 Medium needs 9.9GB, not counting its text encoders.
Is Automatic1111 still worth using?
Not for a new setup. Its master branch hasn't changed since July 2024, and it still asks for Python 3.10.6. Forge Neo keeps the same layout and runs current models.
Is there a newer version than Stable Diffusion 3.5?
No. SD 3.5, released in October 2024, is still Stability AI's latest open image model. The same apps also run other open models, such as Flux, Z-Image Turbo and Krea 2.
Is Stable Diffusion free?
The apps are free, and so are SD 1.5 and SDXL. SD 3.5 is free for personal use and for businesses earning under $1 million a year, which must register with Stability AI for commercial use.
