Running Stable Diffusion locally on windows gives you full privacy, no generation limits, access to thousands of custom models/LoRAs from Civitai, and the ability to use the latest models like SD 3.5, Flux.1, or community fine-tunes — all without paying per image.

As of January 2026, the two most popular and actively maintained ways for most people are:

1. Forge (a performance-optimized fork of the classic Automatic1111 / A1111 WebUI) — easiest for beginners, huge extension ecosystem, familiar tabbed interface.
2. ComfyUI — node-based → more powerful & efficient (especially VRAM), best for complex workflows, video, Flux/SD3.5 native support, but steeper learning curve.

Recommendation for most new users (especially if you’re just starting): → Start with Forge (or plain Automatic1111 if you prefer the original). It’s still the best balance of ease + features in early 2026.

Minimum Hardware Requirements (2026 perspective)

Component	Minimum (usable)	Recommended (fast & modern models)	Notes
GPU	NVIDIA 6 GB VRAM (RTX 2060+)	NVIDIA 12+ GB (RTX 3060/4070+)	AMD possible but slower / more hassle; Intel Arc improving but not ideal yet
RAM	16 GB	32 GB	System RAM, helps with big batches / models
Storage	40–50 GB free	SSD 100+ GB	Models are 4–15 GB each
OS	Windows 10/11, Linux	Windows 11	Mac possible via Draw Things / DiffusionBee but slower on Apple Silicon for big models

If your GPU is weak/old (<6 GB) → many new models won’t run well; consider cloud options like ThinkDiffusion / RunPod instead.

Easiest & Fastest Way: Install Forge (Recommended 2026)

Forge = Automatic1111 but with better speed, lower VRAM use, and better support for 2025–2026 models (SDXL, SD3.5, Flux).

Windows steps (most common):

1. Install prerequisites (one-time)
   • Git → https://git-scm.com/download/win (default settings)
   • Python 3.10.6 → https://www.python.org/downloads/release/python-3106/ (very important: NOT 3.11+ or 3.12 — pick the exact 3.10.6 installer, check “Add Python to PATH”)
2. Download Forge
   • Go to → https://github.com/lllyasviel/stable-diffusion-webui-forge
   • Click green Code button → Download ZIP
   • OR (better, allows easy updates): open Command Prompt and runtextgit clone https://github.com/lllyasviel/stable-diffusion-webui-forge.git
3. Place a model inside
   • Go into the folder → models/Stable-diffusion/
   • Download a good starter model from Civitai.com or Hugging Face:
     • Easy realistic: Juggernaut XL, Realistic Vision, or SD 3.5 Large (~6–12 GB files, usually .safetensors)
     • Put the .safetensors / .ckpt file directly into that folder
4. Launch it
   • Double-click webui-user.bat (or run.bat in some versions)
   • First run → downloads ~10–15 GB of dependencies (Torch, xformers, etc.) → takes 5–25 min depending on internet
   • When finished it shows something like: Running on local URL: http://127.0.0.1:7860
   • Open that link in your browser → you’re in!

Update later: Double-click update.bat (or git pull in the folder if you used git clone).

Alternative: ComfyUI (if you want maximum power & efficiency)

• Better VRAM usage, native Flux/SD3.5 support, reproducible workflows.
• Harder at first (node graph instead of simple tabs).

Quick install:

1. Download from → https://github.com/comfyanonymous/ComfyUI
2. Extract, put models in ComfyUI/models/checkpoints/
3. Double-click run_nvidia_gpu.bat (Windows)
4. Open the local URL it shows.

Many people start with Forge/A1111 → later move to ComfyUI once comfortable.

Quick First Test Prompt (paste into the txt2img tab)

A beautiful cyberpunk city at night, neon lights, rain, flying cars, ultra detailed, cinematic lighting, 8k
Negative prompt: blurry, lowres, bad anatomy, deformed, extra limbs
Sampling method: DPM++ 2M Karras or Euler a
Steps: 25–40
CFG Scale: 7–9
Resolution: 1024×1024 (or 768×1152 etc.)

Stable Diffusion Local Install FAQ (Forge Focus – 2026)

Q1: What exactly is Forge, and why use it instead of the original Automatic1111 (A1111)?

Forge is a heavily optimized fork of Automatic1111’s WebUI. It uses less VRAM, generates images faster (often 2–6× on the same hardware), handles modern models like SD 3.5 / Flux better out-of-the-box, and includes many experimental speed/memory improvements. The interface is almost identical to A1111, so it’s the best “drop-in” upgrade for most people in 2026. If you want maximum customization (node graphs, complex workflows), switch to ComfyUI later.

Q2: Do I really need exactly Python 3.10.6? Newer versions are out—can I use 3.11 or 3.12?

For standard Forge/A1111 installs → yes, stick to Python 3.10.6 (download from the official archive). Many dependencies break on 3.11+ or 3.12 due to Torch/xformers compatibility. Some forks/ComfyUI support 3.11 now, but for the guide’s method → use 3.10.6 and check “Add to PATH” during install.

Q3: The first launch is taking forever / stuck downloading things—what’s happening?

Normal on first run: it downloads ~10–15 GB of PyTorch, xformers, Gradio, etc. (can take 5–40 minutes depending on internet). Let it finish—don’t close the window. If it errors mid-download → close, delete the venv folder inside the Forge directory, and re-run webui-user.bat.

Q4: I double-click webui-user.bat but nothing happens / a black window flashes and closes.

Common causes:

• Python not added to PATH → reinstall Python 3.10.6 with the checkbox ticked.
• Antivirus/firewall blocking → temporarily disable or add exception for the folder.
• Wrong/missing Git → install Git from git-scm.com if you used the ZIP method. Run from Command Prompt instead: navigate to the folder (cd path\to\stable-diffusion-webui-forge) and type webui-user.bat.

Q5: “Out of memory” / CUDA out of memory error even on a decent GPU (e.g., RTX 3060 12GB).

• Lower resolution (start at 768×768 instead of 1024×1024).
• Use fewer steps (20–30).
• Enable –medvram or –lowvram in webui-user.bat (edit the file, add to COMMANDLINE_ARGS).
• Close other GPU-using apps (Chrome with hardware acceleration eats VRAM).
• Forge is already much better at this than plain A1111—SDXL/Flux should run comfortably on 8–12 GB cards.

Q6: No models show up in the dropdown / “No module named…” or loading weights error.

• Place .safetensors files in models\Stable-diffusion\ (not subfolders).
• Restart the UI after adding models.
• If error persists → wrong model format (use .safetensors, not old .ckpt if possible) or corrupted download—re-download from Civitai/Hugging Face.

Q7: Black images / completely blank generations?

Usually VAE mismatch. In Settings → Stable Diffusion → set VAE to “Automatic” or download a matching VAE (e.g., sdxl_vae.safetensors) and place in models\VAE\. Also try a different sampler (DPM++ 2M Karras is reliable).

Q8: How do I update Forge to the latest version?

If you used git clone: Open Command Prompt in the folder → type git pull → restart. If ZIP download: re-download ZIP from GitHub and extract over the old folder (overwrite when asked), then re-run. Many use Stability Matrix (one-click manager) for easier updates—highly recommended if you hate manual git.

Q9: Can I share models/extensions between Forge and ComfyUI / old A1111?

Yes—point them to the same folders. In Forge, edit webui-user.bat to add –forge-ref-a1111-home “C:\path\to\old-a1111” to reuse models/ControlNet/etc. ComfyUI has separate folders but you can symlink or copy.

Q10: I’m on AMD GPU (or Intel Arc)—does this work?

Forge/A1111 NVIDIA-focused → AMD possible via DirectML or ROCm (Linux better for ROCm). Expect slower speeds and more setup. ComfyUI has better AMD support in 2026. If struggling, start with cloud options (RunPod, ThinkDiffusion) or check AMD-specific guides.

Q11: The prompt isn’t following well / bad anatomy— is it broken?

Not install issue—it’s the model. SD 1.5 is old; try modern ones like Juggernaut XL, Pony, SD 3.5 Large, or Flux.1. Add good negative prompts and use CFG 7–9. Install extensions like Adetailer (for faces) via the Extensions tab.

Q12: Where do I get good models / LoRAs safely?

• Civitai.com — biggest community site, ratings, previews.
• Hugging Face (official Stability AI page or search “SD 3.5”). Avoid random sites—stick to these to prevent malware in model files.

Q13: Should I use Forge, ComfyUI, or something else in 2026?

• Beginners/easiest → Forge (this guide).
• Advanced workflows/video/Flux native → ComfyUI.
• One-click manager (handles installs/updates/models) → Stability Matrix. Many users run both Forge + ComfyUI side-by-side.

Q14: My antivirus flags torch or some file as virus—false positive?

Usually yes (PyTorch installers trigger heuristics). Add exception for the entire Forge folder. If paranoid → scan with VirusTotal or download from official GitHub only.

Additional Resources

Stable Diffusion Negative Prompts Guide

Master negative prompting to eliminate artifacts, fix anatomy issues, and generate perfect AI images. Get categorized prompt collections and advanced techniques.

Master Negative Prompts

Stable Diffusion Models Guide

Learn where to download safe models (SD 3.5, Flux.1, custom checkpoints) and how to manage them.

Learn About Models

Troubleshooting Common Issues

Fix black images, out of memory errors, slow generation, and other common problems.

Troubleshooting Guide

Prompting & Advanced Techniques

Master prompt engineering, ControlNet, LoRAs, upscaling, and other advanced features.

Advanced Techniques