Install Stable Diffusion Locally - Private Setup Guide (2026)

The Quick Answer

Key Takeaway - 2026

1. Confirm an NVIDIA GPU with enough VRAM for the model class you want (start by assuming 8 GB minimum for comfortable SDXL-class work).
2. Install Forge (lllyasviel/stable-diffusion-webui-forge) if you want the familiar Gradio WebUI tabs, or Fooocus if you want fewer exposed settings.
3. Drop .safetensors checkpoints into the folders the README names, refresh the UI, generate once at a conservative resolution, then scale up.
4. Or use LocalForge AI if you want a packaged Forge-style install without assembling CUDA choices yourself - one option next to DIY GitHub paths.

Local does not mean lawless. It means you are the operator.

Step 1 - Sanity-check hardware before you download drama

Open Task Manager → Performance → GPU and read your card name and dedicated GPU memory. If you are on 4 GB, plan around smaller checkpoints and 512-pixel workflows until you prove otherwise. If you are on 8 GB, SDXL-class models are realistic with sane settings. If you are on 12 GB or more, you have headroom for heavier stacks - still read each model card, because “Flux-class” weights can stomp even generous cards when you crank resolution.

RAM matters too. 16 GB system RAM is a practical floor; 32 GB reduces paging pain when browsers and the UI both eat memory.

Step 2 - Pick your interface (do not overthink it)

• Forge: Best skeptic’s default when you want extension ecosystems and explicit controls.
• Fooocus: Best when you want fewer tabs and faster “first decent image.”
• ComfyUI: Best when you already know you want node graphs; skip it if you are allergic to JSON workflows on day one.

We are blunt: skip stock AUTOMATIC1111 for new installs unless you have a legacy reason. Compare Forge first; it is the maintained fork people mean when they say “WebUI” in 2026 threads.

Step 3 - Prepare disk layout

Create a folder with no spaces in the path, for example C:\localgen\forge\. Spaces do not always break Python stacks, but they are an avoidable footgun with some scripts.

Budget 30 GB free for the app plus one medium checkpoint; add 20 to 60 GB if you plan to collect models like Pokémon cards.

Step 4 - Install Forge (recommended WebUI-class path)

1. Open the Forge GitHub README and pick the one-click archive that matches your CUDA stack (the README lists paired CUDA and PyTorch builds - do not guess).
2. Extract the archive into your short path folder.
3. Run webui-user.bat on Windows (or the documented shell script on Linux).
4. Wait. First boot can take 10 to 20 minutes while dependencies settle.
5. When the browser opens to http://127.0.0.1:7860, you are in.

If you see CUDA errors, stop and align driver version with the archive you chose. Randomly upgrading PyTorch “because newer is better” is how weekends disappear.

Step 5 - Install Fooocus (alternate minimal path)

1. Clone or download Fooocus from its official GitHub instructions for your OS.
2. Run the documented launch script (run.bat on Windows in common guides).
3. Let first-run downloads finish; do not panic-close when the console looks busy.

Fooocus is not “dumber”; it is opinionated. That is the point.

Step 6 - Place checkpoints and sidecars

For Forge-class layouts, checkpoints usually live under models/Stable-diffusion/ (names vary slightly - trust the README you actually installed). LoRAs land in models/Lora/. VAE files land in models/VAE/ when not baked into the checkpoint.

After copying files, hit the refresh control in the UI. If the model list does not update, you copied into the wrong depth - this is the number one “it is broken” user error.

Step 7 - Generate a boring test image on purpose

Pick a mainstream realistic or anime checkpoint you have a license to use. Set resolution conservatively:

• SD 1.5 class: start near 512×512.
• SDXL class: start near 1024×1024 if VRAM allows; drop to 896×896 if you hit memory errors.

Use 20 to 28 steps and a mainstream sampler (DPM++ 2M Karras is a fine default for many realistic models). If the output is black, you likely have a VAE mismatch - switch to a “baked VAE” build or load the VAE the model card names.

Step 8 - Turn off the cloud brain in your head

Locally, no remote filter runs unless you added something that phones home. That is the privacy win. It is also the responsibility win: archive only what you are allowed to store, do not build non-consensual content, and keep minors out of the pipeline entirely.

Step 9 - Troubleshoot like a skeptic, not a mystic

• CUDA OOM: Lower resolution, enable documented --medvram / low VRAM flags in your launch file, or pick a smaller checkpoint.
• Slow generation: You might be on CPU fallback - verify the UI reports your NVIDIA device.
• Antivirus deletes files: Add exclusions for the install folder; false positives are routine.

Step 10 - Decide what “NSFW” means for your stack

For technical purposes, adult creative work locally is about model choice + prompt + private storage. Commercial use still depends on license. Civitai and Hugging Face are third-party hosts - read each card.

Step 11 - Add LoRAs without wrecking your baseline

LoRAs are small adapter files that steer style or character without replacing the whole checkpoint. Keep them labeled in models/Lora/ with names you understand; “final_final_v7” stops being funny after the third month. Strength values around 0.6 to 0.9 are common starting bands - tune while watching faces and hands, not while chasing forum clout. If skin looks plastic, you probably stacked two style LoRAs that fight each other. Remove one, regenerate, compare.

Step 12 - Back up the stuff that matters

Your time is worth more than disk. Export a text file listing: UI version, GPU driver version, launch flags you use, and the three checkpoints you actually keep. Zip your custom embeddings and favorite small LoRAs separately from giant checkpoints so restores are fast. If you use Comfy later, save one golden workflow JSON beside those notes.

Step 13 - Prove you are not on CPU by accident

After first launch, generate a single image and read the console: it should mention your NVIDIA device. If speeds feel like minutes per image on what should be a gaming GPU, you are often on CPU fallback or a debug path. Fix environment issues before you tune prompts; otherwise you will blame the model for hardware misconfiguration.

When to pause and read the console literally

Skeptics screenshot errors; mystics screenshot vibes. If stderr mentions DLL load failed, copy the full path it prints - half of those traces point at a stale cudnn copied beside the wrong python.exe. Re-run after a reboot once; transient driver reloads after updates cause spooky one-off failures that never reproduce.

Bottom line

You install once carefully, then you iterate for months. Cheating the README buys you reinstall theater.