Easy Local Stable Diffusion Install - Private NSFW Setup (2026)

The Quick Answer

Key Takeaway - May 2026

Install Fooocus v2.5.5 if you want your first image in under 20 minutes with zero decisions. Install Forge (CUDA 12.1 + PyTorch 2.3.1 one-click package) if you want extensions and detailed settings from day one. Both need an NVIDIA GPU with 6+ GB VRAM. Download one SDXL checkpoint (~6.5 GB), generate a test image, then celebrate. Or grab LocalForge AI if you'd rather skip the CUDA pairing step entirely.

What You Need

• GPU: NVIDIA with 6 GB VRAM minimum (RTX 2060 or newer). 8 GB is the comfort zone for SDXL at 1024x1024. AMD works on Linux via ROCm but Windows support is still rough in May 2026.
• Disk space: 15 GB for the app + 7 GB per SDXL checkpoint. Budget 30 GB total if you plan to collect models.
• RAM: 16 GB system RAM recommended. 8 GB works but you'll feel swap pressure with SDXL.
• OS: Windows 10 or 11, 64-bit.
• Time: 20-40 minutes depending on your internet speed (big downloads ahead).

To check your GPU: press Ctrl+Shift+Esc, click Performance, click GPU. You'll see the name and dedicated memory. If it says NVIDIA and shows 6+ GB, you're good.

Step 1 - Download Fooocus (The Easiest Path)

Fooocus is Stable Diffusion with training wheels - and I mean that as a compliment. It picks sane defaults so you can type a prompt and get a good image without touching a single slider.

What to do:

1. Go to github.com/lllyasviel/Fooocus/releases and download the latest Windows package (~2 GB zip).
2. Extract it to a folder without spaces in the path (like C:\AI\Fooocus\).
3. Double-click run.bat inside the extracted folder.
4. Wait for first-run setup - it auto-downloads JuggernautXL (~6.6 GB). Go make coffee.
5. Your browser opens to a single prompt box. Type something simple. Click Generate.

Time to first image: about 15-20 minutes total (mostly download time). The interface is one page with one text box. That's the whole thing.

The tradeoff: fewer controls means fewer ways to customize. That's a feature on night one.

Step 2 - Download Forge Instead (More Control)

Pick Forge if you already know you'll want extensions, img2img tabs, and detailed sampler settings. It's the same familiar WebUI that tutorials reference, but optimized to use 30-50% less VRAM than the old AUTOMATIC1111.

What to do:

1. Go to github.com/lllyasviel/stable-diffusion-webui-forge.
2. Download the one-click package (CUDA 12.1 + PyTorch 2.3.1) from the releases section. The README has a direct link - don't overthink which version, just grab the one it highlights.
3. Extract to a spaceless path (like C:\AI\Forge\).
4. Run webui-user.bat. First boot takes 10-20 minutes while it downloads dependencies.
5. When your browser opens to http://127.0.0.1:7860 - you're done with the hard part.

Important: Forge doesn't bundle a model. You'll need Step 3 before you can generate anything.

Step 3 - Grab One Starter Checkpoint

Don't download ten models before generating a single image. That's how people burn out on night one. Pick one and stick with it until you've proven your setup works.

Good beginner SDXL options (all .safetensors, ~6.5 GB each):

• JuggernautXL - versatile photorealistic style, bundled with Fooocus
• Big Lust v1.6 - photorealistic NSFW merge, 74K+ downloads on Civitai
• Photo Momentum v2.0 - quality photorealism without needing extra LoRAs

Download from Civitai or Hugging Face. Always grab the .safetensors version (safer format). Read the model's license before you get attached to the outputs - some restrict commercial use.

For Forge: drop the file into models/Stable-diffusion/ inside your Forge folder. Hit the refresh button in the UI's model dropdown. If it doesn't appear, you're probably one folder too deep.

For Fooocus: it already has JuggernautXL. If you want a different model, check Fooocus docs for the correct models folder path.

Step 4 - Generate a Boring First Image

Resist the urge to write a complex prompt. Your first generation is a hardware test, not art.

Try this prompt: "a white cup on a wooden table, soft natural light, simple background"

Settings that won't fight you:

• Resolution: 1024x1024 for SDXL models (768x768 if you have exactly 6 GB VRAM)
• Steps: 25 (safe middle ground for most samplers)
• CFG scale: 7 (Forge default, works fine)

If you get an image of a cup - congratulations, your stack works. If you get black output or neon garbage, that's usually a VAE mismatch. Check the model page for "recommended VAE" notes.

Step 5 - Your First Real Prompt

Now that you know the plumbing works, try something you actually want to generate. A few tips that'll save you frustration:

• Lock your seed while learning. Random seeds make it impossible to tell what your prompt changes actually did.
• Change one thing at a time. Don't adjust prompt, steps, CFG, and resolution all at once.
• Add detail gradually. Start with "woman, red dress, studio lighting" before adding 50 modifiers.
• Use the model's trigger words if the Civitai page lists them. Some checkpoints need specific keywords to activate their style.

Step 6 - Build Good Habits Early

• Save your first good image along with the settings that made it (model name, resolution, sampler, steps, CFG, seed). You'll thank yourself later.
• Take breaks. Local gen is addictive because iterations feel free. Set a 45-minute timer.
• Strip metadata before posting anywhere. Some tools embed your full prompt in the PNG file. Use the "save without metadata" option if your UI has one.
• One LoRA at a time. They're fun but they stack weirdly. Add them individually so you know what each one does.

Verify It Works

You know your setup is solid when:

• The UI loads without Python errors in the terminal
• Your model appears in the model dropdown
• A simple prompt produces a coherent image (doesn't need to be pretty)
• Generation takes 8-20 seconds on 8 GB VRAM at 1024x1024 (not minutes - that means CPU fallback)

If generation takes over 60 seconds for a single image, check the terminal window. It'll usually say "running on CPU" somewhere. That means CUDA isn't working - revisit the one-click package choice.

Troubleshooting

• Out of memory (OOM): Lower resolution to 768x768, close Chrome tabs (browsers eat VRAM too), or add --medvram to your launch arguments in webui-user.bat.
• CUDA error on launch: You grabbed the wrong one-click package. Re-read the Forge README table and match your GPU generation. RTX 30/40 series = CUDA 12.1 package.
• Black or neon output: VAE mismatch. Download the VAE file listed on your model's page, place it in models/VAE/, and select it in Settings.
• Model not showing in dropdown: Wrong folder depth. The .safetensors file should be directly inside models/Stable-diffusion/, not in a subfolder.
• Generation takes minutes: You're running on CPU. Check that NVIDIA drivers are updated and that the terminal shows "Using CUDA" or similar on startup.

What to Do Next

• Learn Forge's full install options: Forge Installation Guide covers advanced launch flags, multi-GPU setup, and extension management.
• Check your hardware fits: Hardware Requirements Guide has real VRAM measurements for every model type.
• Explore the free setup stack: Free Local NSFW Setup covers the full free toolchain beyond just the UI.

Bottom Line

The whole path is: check GPU, download Fooocus or Forge, grab one model, generate a test image. That's four real actions. Everything else - LoRAs, upscalers, custom VAEs - is optional until you've proven the base works. Don't let setup become the hobby. The hobby is making images.