Cloud image generators charge per image and filter your prompts. Running Stable Diffusion locally is free, private, and unlimited: your prompts never leave your machine, there are no rate limits, and you can use any of the thousands of community models from sites like Civitai. Here is how to set it up on a Windows, Linux, or Mac machine in 2026.
What you need first
Stable Diffusion runs best on a discrete GPU. The realistic minimums in 2026 are:
- 4 GB VRAM (e.g. GTX 1060): works with SD 1.5 models using low-VRAM flags
- 6 GB VRAM (e.g. RTX 3060 laptop): comfortable with SD 1.5, tight for SDXL
- 8 GB+ VRAM (e.g. RTX 3060 Ti / 4060 Ti): SDXL at 1024x1024 runs smoothly
- Apple Silicon Mac: works via ComfyUI, slower but usable for SD 1.5/SDXL
You also need Python 3.10 or 3.11 (3.12 still breaks some extensions), Git, and roughly 20–30 GB of free disk space — checkpoints are 2–7 GB each and you will collect several.
Step 1: Pick your interface — ComfyUI or Forge
Two frontends dominate local Stable Diffusion in 2026, and choosing well saves you a reinstall later.
ComfyUI is a node-based workflow canvas: you wire generation steps together visually, save workflows as JSON or PNG metadata, and reuse them exactly. It is the fastest option, the best for animation/video extensions, and the one most new tooling targets. It listens on http://127.0.0.1:8188.
Stable Diffusion WebUI Forge (lllyasviel's fork, continued by the community as Forge Neo) is an optimized take on the classic AUTOMATIC1111 interface with txt2img, img2img, inpainting, LoRA, and ControlNet in a familiar form layout. The original AUTOMATIC1111 has had no release since v1.10.1 (July 2024), so Forge is the maintained choice for that style of UI.
Rule of thumb: pick ComfyUI if you want maximum speed, reproducibility, and access to cutting-edge workflows. Pick Forge if you prefer a simple prompt-box interface and do not want to learn nodes. You can install both later — they can share one model folder.
Step 2: Install with Stability Matrix (easiest path)
The least error-prone installer in 2026 is Stability Matrix, a free launcher that installs and updates ComfyUI, Forge, and SwarmUI with its own bundled Python and Git, so nothing conflicts with your system install.
1. Download the Stability Matrix release zip from its GitHub releases page and extract it anywhere — it is portable, with no installer.
2. Launch the app, click Add Package, and choose ComfyUI (or Forge).
3. Let it finish downloading dependencies. You now have a launch button per package, a built-in model browser that downloads checkpoints directly from Civitai, and one shared models folder all packages can use.
Prefer the manual route? For ComfyUI, clone the repo and run the portable install:
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
# Windows (NVIDIA): run install + start via the provided .bat
# Linux: pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
python main.pyThen open http://127.0.0.1:8188 in your browser.
Step 3: Download a checkpoint model
The interface is useless without a model. Two safe starting points:
- SDXL Base 1.0 (Stability AI, on Hugging Face): the best general model for photorealism and art at 1024x1024. About 6.9 GB as a safetensors file.
- SD 1.5 checkpoints (e.g. community merges on Civitai): lighter (4.3 GB), fast on 4–6 GB VRAM, and the largest ecosystem of styles and LoRAs.
Download the .safetensors file (never run .ckpt files from untrusted sources — safetensors cannot execute malicious code) and place it in the checkpoints folder inside your install's models directory. In Stability Matrix, the model browser does this for you automatically.
Step 4: Generate your first image
In Forge: open the txt2img tab, type a prompt, set size to 1024x1024 for SDXL (512x512 for SD 1.5), and click Generate. Add a negative prompt like "blurry, low quality, deformed" to raise the floor on quality.
In ComfyUI: load the default text-to-image workflow, select your checkpoint in the "Load Checkpoint" node, type positive and negative prompts, set width/height to match the model's native resolution, and press Queue Prompt. Your first image should appear in 2–15 seconds on a mid-range GPU.
The single biggest quality lever is native resolution: SDXL was trained at 1024x1024, SD 1.5 at 512x512. Stray far from those and faces and hands fall apart. Upscale afterwards instead — the built-in ESRGAN/Real-ESRGAN upscalers or a tiled-diffusion workflow can take you to 4K.
Step 5: Low-VRAM optimisation flags
On a 4–6 GB card, add these to your launch arguments:
--medvram: splits the model across GPU/CPU memory (6 GB cards)--lowvram: maximum memory saving, slower (4 GB cards)--xformers: faster attention with less memory on NVIDIA cards
Expect roughly 30–60 seconds per image on a 4 GB card with these flags, versus a few seconds on an 8 GB card. CPU-only generation works but takes 5–10 minutes per image — fine for experimenting, not for batch work.
Step 6: Expand with LoRAs and ControlNet
Once generation works, two extensions change everything:
- LoRA (Low-Rank Adaptation): small 50–200 MB files that add a style, character, or concept to your checkpoint. Load them through the LoRA loader node in ComfyUI or the LoRA tab in Forge. Civitai hosts tens of thousands, most free.
- ControlNet: steers generation with depth maps, poses, or sketches. In ComfyUI this is a set of nodes plus a ControlNet model file; in Forge it installs as an extension.
Troubleshooting common failures
- "Torch is not able to use GPU": your PyTorch build does not match your CUDA version. Reinstall the correct CUDA-enabled torch build for your GPU.
- Out-of-memory errors: lower the resolution, close other GPU apps, or add
--medvram/--lowvram. - Python 3.12 extension errors: drop to Python 3.10 or 3.11; many extensions still pin older versions.
- Slow first launch: the first run downloads extra dependencies and can take 10+ minutes — that is normal, not a crash.