Beginner 11 minImage

Stable Diffusion locally: installation and models

Direct response

Stable Diffusion is a family of open-weight image models, not an application: you run it in a local interface. To get started, use SDXL (6.94 GB file, 8 GB of comfortable VRAM) in Fooocus, ComfyUI Desktop, or, on Mac, Draw Things. SD 1.5 works for smaller cards; SD 3.5, announced on October 22, 2024, is for cards with 12 GB or more. Newer families, such as Flux or Z-Image, mainly run through ComfyUI.

Stability AI announced the public release of Stable Diffusion on August 22, 2022: a model that turns text into images on your own computer, without an account or volume limit. As of September 28, 2026, the name covers three generations of models, four or five possible interfaces, and licenses that differ substantially. This guide helps you choose a version and an interface, install everything on Windows, Mac, or Linux, and understand what you’re allowed to do with the images.

By Mohamed Meguedmi·Update 2026-09-29·Tested on Windows, macOS, and Linux

#What is Stable Diffusion?

Stable Diffusion is a latent diffusion model. It starts with an image of pure noise and denoises it step by step, guided by your text, until a coherent image emerges. The computation takes place in a compressed space called latent space, which explains why a consumer graphics card is sufficient where the first diffusion models required a computing center.

What made it famous wasn’t its raw quality but its openness. When it was released in 2022, Stability AI chose a CreativeML OpenRAIL-M license, which it describes as “permissive” and allowing both commercial and noncommercial use: the weights can be downloaded, modified, and retrained. Thousands of specialized variants exist. Stable Diffusion isn’t an application: it’s a model file that you run in the interface of your choice.

The name creates a vocabulary trap. The reference SD 1.5 repository on Hugging Face is now presented as a mirror of the old RunwayML repository, deprecated and unrelated to that company. And since 2024, the most discussed open image models have come from other labs: Flux, Qwen Image, or Z-Image. They run in the same interfaces, but they are not called Stable Diffusion.

#The versions, and which one to choose

The Local Image AI Kit

AI images and videos on your own machine, no subscription and no credits: ComfyUI, Flux, Z-Image and Wan 2.2 with ready-to-load workflows, VRAM tiers, LoRA training and the legal frame.

  • Lifetime online access
  • PDF + files
  • Lifetime updates
The Local Image AI Kit →or all kits for life — $49 →
The three generations of Stable Diffusion locally, according to Stability AI and the Hugging Face model cards
VersionOutputMain fileNative resolutionComfortable VRAMKey takeaway
SD 1.520224.27 GB (v1-5-pruned-emaonly)512 × 5124 to 6 GBLightweight, with a huge library of variants and LoRAs. Details and hands could be better.
SDXL 1.0July 20236.94 GB (sd_xl_base_1.0)1024 × 10248 GB: Stability AI claims it runs properly on 8 GB cardsThe right starting point: significantly better quality, a mature ecosystem.
SD 3.5October 22, 2024Medium: 2.5 billion parameters; Large: 8.1 billion; Large Turbo: 4-step versionMedium: 0.25 to 2 megapixels; Large: 1 megapixelMedium: 9.9 GB excluding text encoders (Stability AI). Large: more than 16 GB in 16-bit (our calculation)Better prompt adherence. Different license; see below.

Two details can help you avoid purchasing or downloading mistakes. First, SD 1.5 comes in two files: v1-5-pruned-emaonly.safetensors (4.27 GB), which the model card describes as using less VRAM and being suitable for inference, and v1-5-pruned.safetensors (7.7 GB), intended for retraining. For image generation, choose the first one. Next, SD 3.5 Large, with its 8.1 billion parameters, should not be evaluated based on the Medium model card: do not count on it running on an 8 GB card.

If you're starting with an 8 GB card, begin with SDXL: it offers the best balance of quality, speed, and available documentation. With 12 to 16 GB, also look at recent ComfyUI models: its README announces native support for Flux.1, Flux.2, Qwen Image, and Z-Image.

#The required hardware

What works with your GPU · cards from the QuelLLM hardware database
Your hardwareWhat works well
Card NVIDIA 4 to 6 GB (GTX 1660, RTX 2060, RTX 3050)SD 1.5. Fooocus, under SDXL, lists 4 GB of VRAM and 8 GB of RAM as the minimum, relying on Windows virtual memory: slow, but possible.
NVIDIA card 8 GB (RTX 3060 Ti, RTX 4060, RTX 5060)SDXL comfortably, threshold announced by Stability AI
NVIDIA card, 12 to 16 GB (RTX 3060 12 GB, RTX 4070, RTX 5070 Ti)SDXL with LoRA and upscaling, SD 3.5 Medium, Z-Image Turbo (16 GB according to its listing), Flux in FP8
Recent AMD card (RX 7000, RX 9000)Good on Linux with ROCm; on Windows, official ROCm support in ComfyUI Desktop since January 2026
Mac Apple SiliconAny M-series chip: unified memory serves as VRAM. Slower than a NVIDIA card in the same performance class.
No graphics cardTechnically possible, but much slower: reserve it for testing

Plan for disk space too. A single SDXL file weighs 6,94 GB, and collections of LoRAs and variants grow quickly. Fooocus recommends having at least 40 GB free on each disk if you encounter the “RuntimeError: CPUAllocator” error.

#Which interface should you install?

Interfaces to know, as of September 28, 2026
InterfaceWho it's forStrengthsLimitations
FooocusBeginnersA text field, a button, carefully chosen default settings; fewer than three clicks from download to the first image, according to its READMESDXL only: the project has limited long-term support, with bug fixes only
Forge (lllyasviel's repository)Tabbed-interface usersThe more memory-efficient AUTOMATIC1111 interface, with Flux supportNo commits to the main branch since June 2025; development continues through forks such as Forge Neo
AUTOMATIC1111Those following older tutorialsThe legacy interface, the best documentedLatest version released in February 2025, Flux not supported
ComfyUIAdvanced users and curious mindsNative support for recent models, reproducible workflows, very active projectLearning curve
Draw ThingsMac, iPhone, iPadNative application, free edition that computes locally and offline, on-device LoRA trainingApple ecosystem only; paid cloud offerings are available as an option, but avoid them if you want to stay 100% local
!
Fooocus: beware of fake sites
The Fooocus README warns that many fake sites appear when you search for its name, and lists fooocus.com, fooocus.net, fooocus.ai, and fooocus.org as all fake. The project's only official source is its GitHub repository. The same advice applies to other software: start from the repository, not an ad.

#The shortest path per system

  1. 01
    Windows with a NVIDIA card
    Fastest: Fooocus. Download the archive from the official GitHub page, extract it, and run run.bat. On the first launch, the models are downloaded to Fooocus\models\checkpoints. For a tool that keeps up with new models, install ComfyUI Desktop instead, recommended for new users by the ComfyUI README.
  2. 02
    RTX 50 card: check PyTorch
    Forge’s all-in-one archives list CUDA 12.1 and PyTorch 2.3.1 as the recommended versions. However, RTX 50 cards require a recent PyTorch version: an AUTOMATIC1111 contributor mentions PyTorch 2.7.0 as the first official version to support them. On an RTX 50, prefer ComfyUI, whose documentation requires CUDA 13.0 for recent NVIDIA cards.
  3. 03
    Mac Apple Silicon
    Install Draw Things from the App Store: its free edition runs locally and offline, with no command line. Do not enable its cloud options if you want nothing to leave your Mac. ComfyUI Desktop is also available for macOS on Apple Silicon.
  4. 04
    Linux, or an AMD card
    Manual ComfyUI installation: clone the repository, set up a Python environment, and install the CUDA or ROCm version of PyTorch depending on the GPU. The detailed procedure is in our ComfyUI guide. Fooocus also offers a Linux installation via Anaconda.
  5. 05
    Verify that the GPU is working
    During generation, nvidia-smi should show the VRAM in use and the GPU near 100%. If the image takes several minutes, the interface is running on the CPU: it is almost always an unsuitable PyTorch version.

#Where to find the models, and which ones to avoid

Two sources dominate. Hugging Face hosts official models from research labs. Civitai brings together community variants: fine-tuned models, LoRAs, and styles. A LoRA is a small file that adds a style or subject to a base model without replacing it. It must match the right family: an SDXL LoRA does not work on SD 1.5.

!
Three precautions
Download only .safetensors files: according to the format's repository, it stores tensors “safely (as opposed to pickle),” whereas the older .ckpt format can execute code when opened. Then read the details for each community model, because many are trained on content whose commercial use is prohibited or legally uncertain. Finally, choose the right file: for SD 1.5, the “emaonly” variant is for generation, while the other is for retraining.

#Writing a prompt that works

Describe; don't issue commands
These models don't follow instructions. “Portrait of an older woman, window light, 85 mm lens, shallow depth of field” works; “take a beautiful photo of me” doesn't.
Write in English
The SD 1.5 model card says “Language(s): English” and specifies a frozen CLIP ViT-L/14 text encoder. French is understood, but less precisely.
Subject, then framework, then style
Order matters: what comes first carries the most weight.
The negative prompt
It lists what you don't want to see. It's especially useful for SD 1.5 and SDXL; check a recent model's card to see whether it takes this into account.
Keep the seed
Same seed, same settings, same image. To compare two prompts, fix the seed and change only one thing at a time.

#Licenses and commercial use

Licenses for the three generations, based on Stability AI publications and Hugging Face model cards
ModelLicenseCommercial use of images
SD 1.5CreativeML OpenRAIL-MAllowed: Stability AI describes it as permissive for commercial and noncommercial uses, with ethical-use restrictions listed in the license
SDXL 1.0CreativeML OpenRAIL++-MAllowed, same principles
SD 3.5Stability AI Community LicenseFree for noncommercial use and for commercial use up to 1 million dollars in annual revenue; beyond that, contact Stability AI for an enterprise license. You retain ownership of the generated images.

The model license does not settle everything. Copyright in a generated image and the use of a real person's likeness are governed by the laws of your country, not the software license. This guide describes the licenses published by their authors and does not constitute legal advice.

#FAQ

FAQ
Is Stable Diffusion free?+
Yes. The models download for free, and the interfaces that run them are open source. Locally, you pay neither subscriptions nor credits, only for your hardware and electricity. One licensing exception: SD 3.5 is free for commercial use up to 1 million dollars in annual revenue, with an enterprise license required beyond that. Paid online services exist, but they are not necessary.
Can Stable Diffusion be used without a graphics card?+
Technically, yes, but generation becomes significantly slower. For practical use, plan on an NVIDIA card with 8 GB for SDXL, the threshold Stability AI states for proper operation, or a Mac with Apple Silicon. Fooocus can go as low as 4 GB of VRAM and 8 GB of RAM on Windows, at the cost of speed. Without a GPU, limit testing to a few images.
Which version should you choose: SD 1.5, SDXL, or SD 3.5?+
SDXL for a first try: 8 GB of VRAM is enough, and the ecosystem is mature. SD 1.5 if your card has 4 to 6 GB, with more modest images. SD 3.5 Medium requires 9.9 GB of VRAM excluding text encoders, according to Stability AI: reserve it for cards with 12 GB or more, or when prompt adherence matters.
Do my images remain private?+
Yes, if you stay local: the prompt and image never leave your machine. Draw Things says it works offline in its free edition, and ComfyUI’s README states that its core downloads nothing unless requested. The only possible leaks come from online options you enable or a malicious extension: install from official repositories.
Stable Diffusion or Flux?+
Stable Diffusion remains lighter and better documented: SDXL fits in 8 GB. Flux.1 dev weighs 23.8 GB at full precision, and Comfy-Org’s FP8 checkpoint is 17.2 GB, making them better suited to 12–16 GB cards and above. Choose Flux for prompt fidelity if your card can handle it; otherwise choose SDXL, with its larger LoRA ecosystem.
Can you use Stable Diffusion on Mac?+
Yes, on Apple Silicon chips. Draw Things is an app whose free edition runs offline on Mac, iPhone, and iPad. ComfyUI Desktop also exists for macOS, and the AUTOMATIC1111 wiki documents a Mac installation where training remains very slow. NVIDIA cards are still faster at a comparable tier, but a Mac is enough to learn.
Can you sell generated images?+
Often yes, but it depends on the model. SD 1.5 and SDXL, under OpenRAIL licenses, allow commercial use with usage restrictions. SD 3.5 is free under 1 million dollars in annual revenue. Also read the profiles for the added community models and LoRAs, and remember that copyright on a generated image is governed by your country.
Did this guide help you?

Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.