LocalForge AILocalForge AI

Fooocus Stable Diffusion Installation - 2026 Guide

Fooocus is the tool we recommend when someone says they want Stable Diffusion results without learning every sampler name first. It keeps the surface area small, pushes sensible defaults, and gets you to a decent image before your coffee cools. This guide is advisory but not hand-holding: you still need a NVIDIA GPU for a smooth ride, enough disk for checkpoints, and the discipline to read the official Fooocus GitHub instructions for the branch you actually download. We cover Windows-first steps because that is the path most readers take; macOS and Linux users should translate paths and rely on the same principles - short folder names, patient first boot, and no antivirus theater. Adult creative use is the same technical stack as any other local generation: you pick legal models, you store outputs privately, and you skip anything that harms real people. If you want Forge-style tabs later, you can add a second install; Fooocus plays nicely as the “fast sketch” machine in that pairing.

Who wrote this

Written by the LocalForge AI team. LocalForge AI is our own product, an offline image generator, so we have a stake in this topic. We still recommend Fooocus here because it is free, and we wrote this guide for people who want to install it themselves.

The Quick Answer

Key Takeaway - 2026

  1. Create a short path folder (no spaces) on a drive with tens of gigabytes free.
  2. Clone or download Fooocus from the official GitHub repo and follow its Windows quick start (run.bat in common instructions).
  3. Let first-run downloads finish; the wait depends mostly on your disk and connection (see Step 6).
  4. Drop new checkpoints into the paths the README documents; refresh or restart as instructed.
  5. Generate at a conservative resolution first, then raise it in small steps.
  6. Or use LocalForge AI if you decide you want a packaged Forge route instead. LocalForge AI is our product; we recommend Fooocus here because it is free, and the product is the alternative for people who would rather not install anything.

Check which version is current

This guide follows the current official Fooocus release on GitHub, so check the repo's releases page and its release discussion (the archived one is titled "v2.5.0 - incl. how to update") to see what is current on the day you install. In Step 4 you download a release archive; use the Windows archive that the release page lists, and write its exact file name down as Step 19 describes. To move to a newer release later, follow the update path in Step 14.

Step 1 - Decide Fooocus is the right “advisor pick”

Choose Fooocus when you want fewer tabs and faster iteration on prompts. Pick Forge instead when you already know you want extensions, XYZ grids, and script hooks. Pick ComfyUI when you want node graphs. If you are unsure, Fooocus is the least punitive first night.

Step 2 - Check GPU memory like an adult

Open your GPU panel and read dedicated VRAM. Use these as planning bands, not promises:

  • 4 GB: Tight. Expect smaller models and lower resolution unless you enjoy error messages.
  • 8 GB: Comfortable for many SDXL-class workflows with Fooocus defaults.
  • 12 GB+: More room for heavier checkpoints and less time fighting settings.

These bands are guidance, not our measurements. Your own card, driver and checkpoint decide the real limit, so treat the first test image in Step 7 as the real check.

System RAM: treat 16 GB as a practical minimum; 32 GB is calmer when browsers stay open.

No NVIDIA GPU? CPU and AMD

Fooocus can run on CPU, but it is far slower than on an NVIDIA card. We did not test CPU-only speed, so we give no figure for it. If you try it anyway, use a smaller resolution and be patient with every image. AMD cards are not the route this guide walks through, so check the official README for the current notes on your card. If waiting minutes per image is a dealbreaker, move to a different route, such as a packaged offline generator like LocalForge AI, or a machine with an NVIDIA GPU.

Mac and Apple Silicon

We wrote this guide for Windows with an NVIDIA GPU. We did not test Fooocus on an Apple Silicon Mac, so we cannot tell you it is a smooth route. Apple chips do not use CUDA, so expect the setup to differ from the Windows steps here, and treat the official README's macOS instructions as the source of truth. Linux users should likewise follow the README and translate the paths and launch commands in this guide.

Step 3 - Prepare folders

Example: C:\localgen\fooocus\. Avoid C:\Program Files\ for Python stacks - permission friction is real.

Step 4 - Get the code

Use git clone if you plan to pull updates, or download a release archive if you want a frozen snapshot. Either path is fine; pick based on whether you like terminal hygiene or zip hygiene.

Step 5 - First launch on Windows

  1. Open the folder in Explorer.
  2. Run run.bat (name may vary slightly by branch - match the file the README names).
  3. Read the console instead of closing it in panic. Warnings are data.

If Windows SmartScreen complains, verify you are in the official repo, then allow. Random repacks from file hosts are how wallets get lighter.

Step 6 - Wait through dependency setup

First boot downloads Python pieces, models, or components depending on the exact Fooocus build. Expect a long first boot, longer on slow disks, and leave the console open until it finishes. Plug into Ethernet if Wi-Fi is flaky - partial downloads corrupt patience.

Step 7 - Open the UI and click one boring test

When the UI appears, generate a simple subject with neutral wording. You are validating install health, not art direction. Write down the prompt, seed, preset and resolution of this first image, so you can repeat it exactly if a later run misbehaves. If generation completes in a sane time window on your GPU tier, you are done with engineering and ready for creative work.

Step 8 - Add checkpoints like a librarian

Download .safetensors from sources you trust - official model pages, Hugging Face, or Civitai cards with clear licenses. Place files in the directory the README specifies (often a models subtree). Too many beginners copy files one folder too deep; if the UI cannot see the model, you misplaced it.

Step 9 - Resolution discipline

Start below your dream resolution, then increase:

  • For many SDXL-class models, a smaller first resolution is a safer first success than jumping straight to the model's full native size.
  • If you hit CUDA OOM, lower the resolution one step at a time until stability returns.

Step 10 - Styles, presets, and not over-stacking

Fooocus gives you style presets because most people do not want to hand-tune fifty sliders. If results look muddy, remove a style layer before you blame the checkpoint. Advisors recommend one big change per iteration so you know what moved.

Step 11 - Privacy and “uncensored” wording

Locally, you choose weights and prompts; there is no cloud classifier in the default loop. That is privacy and creative control. It is not a license to break law, steal likeness without consent, or sexualize minors. Keep your storage encrypted at rest if your threat model says so - BitLocker on Windows, sensible file permissions, separate user accounts on shared PCs.

Step 12 - When to graduate to Forge or Comfy

If you need LoRA strength per block, ControlNet stacks, or batch automation, you will outgrow minimal UIs. Install Forge alongside Fooocus; keep Fooocus for quick comp ideas. Symlink shared checkpoints only after you read both README path sections.

Step 13 - Troubleshooting cheatsheet

  • Model not listed: Wrong folder depth or wrong extension; re-read README paths for your exact commit.
  • Black image: VAE mismatch; try a baked-VAE checkpoint build or add the VAE named on the model card.
  • Minutes per image: Confirm you are on GPU, not CPU fallback.
  • Antivirus nuked files: Exclusions for the install folder.

If you hit CUDA out of memory, lower the resolution as in Step 9. Whatever fails, copy the exact message from the console before you change anything; it is the first thing a support thread needs (see Step 18).

Step 14 - Update without breaking your week

If you cloned with git, git pull plus a documented dependency refresh beats deleting everything. If you use a frozen zip, snapshot your models downloads before replacing the app folder. Never update ten minutes before a deadline - GPU stacks love irony.

Step 15 - Prompt habits that save VRAM

Long prompts do not always cost more VRAM, but chaotic prompts cost time because you iterate blindly. Keep a scratch file with three working bases - portrait, landscape, stylized - and branch from those. Change one clause per experiment. Your future self reads logs better when yesterday-you was disciplined.

Step 16 - Disk hygiene for people who hoard checkpoints

Duplicate downloads are the silent tax. Name files with model version + date when mirrors confuse you. Move cold checkpoints to an external drive and symlink only if you understand Windows symlink tools; otherwise keep one “hot” library on NVMe and accept manual moves.

Step 17 - Thermal reality on laptops

Thin laptops throttle. If your first hour is fast and the second hour crawls, check temperature and power mode. A 15 W sustained cap turns a gaming SKU into a space heater with regrets. Use a stand, clean dust annually, and expect lower stable clocks than desktop counterparts.

Step 18 - When to ask for help without shame

Paste three things in support threads: GPU model, VRAM size, and the first CUDA error block. Skip the two-paragraph backstory. Communities answer faster when you behave like an engineer, not a novelist.

Step 19 - Log retention for future you

Keep a dated .txt next to the install with the exact commit hash or release zip name you used. When something regresses after an update, you will know which baseline to diff against instead of guessing which Tuesday broke VAE loading. Your brain is not a changelog.

Bottom line

Fooocus rewards patience on first boot and discipline on folder hygiene. Nail those, and the rest is prompting skill.

FAQ

Does Fooocus run on AMD GPUs? +
Some community paths exist with extra friction. For predictable interactive speeds in 2026, NVIDIA remains the default recommendation.
Can I use commercial models? +
Only if the model license allows your use case. Read each card; local generation does not bypass copyright.
Is Fooocus outdated compared to forks? +
Forks like RuinedFooocus target different feature mixes. Pick based on the README of the repo you actually install, not forum branding.
Do I need a cloud account? +
Core local generation does not require cloud accounts for weights you already downloaded. Some auxiliary downloaders may differ - read what you run.