AiBOX.im

Fooocus

Dead-simple AI image gen (one-click, fewer knobs than A1111)

— views · — installs
WindowsmacOS Difficulty ●●●●○ stable v2.5.5
Fooocus icon

Fooocus · icon from the registry

System requirements

No extra runtime needed

This app ships with everything it needs — no Node.js, Python or other runtimes to install.

Platforms

WindowsmacOS

Dead-simple AI image gen (one-click, fewer knobs than A1111)

Registry data updated: 2026-11-16

Config & data

Config file written to

~/SillyTavern/data/default-user/settings.json

  • ~/Fooocus
  • ~/Fooocus

Everything stays on your own machine — nothing is uploaded. To uninstall, delete these folders.

Install in three steps

  1. 01

    Install AiBOX

    One portable file for Windows (macOS coming). Double-click and it runs — no installer, no administrator rights.

    Go to download →
  2. 02

    Pick Fooocus

    Open the AiBOX app catalog, find Fooocus and press Install. AiBOX checks what your machine already has and installs only what is missing.

  3. 03

    Follow the wizard

    It downloads with a checksum check, writes the config and fills in your API key if the app needs one. You just click.

    ✓ Installed until it works.

Prefer to do it by hand? The full written guide is right below. Read the guide ↓

Install guide

The same steps in written form — follow along at your own pace.

One line: Fooocus is one-click AI image generation — Midjourney's ease of use with Stable Diffusion's free, offline, local-first model: write a prompt, get a good image, no parameter PhD required (4GB VRAM is enough).

What you get

  • Dead-simple generation: the project's own goal is "fewer than 3 clicks between downloading and your first image", and prompts work even when they're sloppy (built-in GPT-2 prompt expansion)
  • Fully offline and open source: once models are local, it works with the network unplugged and nothing leaves your machine
  • Depth when you need it: style packs, presets, img2img / inpaint / upscale / face swap, plus SDXL models and LoRAs straight from Civitai

Fooocus or A1111? (read this first)

You wantPick
Open it and generate, zero parameters to learnFooocus (this page)
Plugins (ControlNet etc.), per-parameter control, batch automationStable Diffusion WebUI (A1111)
Models & prompt syntaxThey overlap: Fooocus uses A1111's reweighting algorithm, so (word:1.5) and embedding syntax carry over unchanged

Before you start (3 minutes)

NeedNotes
Windows 10/11 or macOS 12+Linux works too; no admin rights
Git + Python 3.10Source route requires Python 3.10; Windows users can skip Python with the official one-click bundle
GPUMinimum 4GB NVIDIA VRAM + 8GB RAM (pre-Pascal GPUs unsupported); AMD goes through DirectML (~3× slower), Mac/CPU much slower
DiskFirst launch auto-downloads ~2.5GB of PyTorch plus multi-GB SDXL weights — keep 20GB+ free

Install

Route A: install with AiBOX (recommended)

  1. Open AiBOX → app directory → Fooocus → Install
  2. It downloads the official tag archive (hash-verified, CN traffic switches gateway automatically) and clones into ~/Fooocus
  3. A user-level venv is created and requirements_versions.txt is installed — no admin rights, system Python untouched

Route B: manual install (official commands)

git clone --depth 1 -b v2.5.5 https://github.com/lllyasviel/Fooocus.git ~/Fooocus
cd ~/Fooocus
python -m venv venv
./venv/bin/pip install -r requirements_versions.txt    # Windows: .\venv\Scripts\pip
./venv/bin/python entry_point.py                      # the official Windows bundle uses run.bat

On first launch models are downloaded into Fooocus/models/checkpoints (pre-place them to skip the wait).

Configuration & verification

Generation parameters — the five that matter

  1. Performance: Speed (30 steps, default) / Quality (60 steps, finer) / Extreme Speed and Lightning (8 or 4 steps with their LoRAs, very fast). Start on Speed, switch to Extreme Speed when rushing, Quality when polishing
  2. Style: tick style templates from the sdxl_styles folder — keep Fooocus V2 (prompt expansion) on by default; put what you *don't* want into Advanced → Negative Prompt
  3. Aspect Ratios and Image Number are self-explanatory; Guidance / Sharpness live under Advanced → Advanced (lower = more obedient, higher = more stylized)
  4. Prompt weighting: same as A1111 — (sunset:1.4) to emphasize; anti-prompts go in the Negative Prompt
  5. Input Image: drop an image to get Upscale / Variation / Inpaint / Outpaint / Image Prompt (including FaceSwap)

Presets and models (style packs)

  • Switch the preset at the top: default (all-round, juggernautXL_v8Rundiffusion), realistic (realisticStockPhoto_v20), anime (animaPencilXL_v500) — switching auto-downloads the matching model, or block that with --disable-preset-download and place models manually
  • From the CLI: --preset anime. Your own checkpoints (Civitai SDXL models) and LoRAs go into models/checkpoints / models/loras, then get picked in the UI

Offline mode

Once dependencies and weights are downloaded, Fooocus is offline software — it generates with the network unplugged. On a flaky connection use --hf-mirror https://hf-mirror.com to fetch models from a CN mirror, and --disable-preset-download to stop any automatic downloading altogether. Heads-up: the first inpainting run also pulls a 1.28GB control model, so do one warm-up pass while online.

./venv/bin/python entry_point.py --preset realistic --port 7860

Other handy flags: --listen (LAN access), --language (UI language), --theme dark.

What "success" looks like

  1. The terminal prints http://127.0.0.1:7860 — open it in a browser
  2. Type any prompt ("a cat in a garden") and hit Generate
  3. An image appears within seconds to a minute, with its generation info viewable = you're set

FAQ

  • First launch is painfully slow: expected — it's downloading PyTorch and models; watch the terminal
  • MetadataIncompleteBuffer / PytorchStreamReader: the model file is corrupt; delete and download again
  • Out of VRAM / RuntimeError: CPUAllocator: check the 4GB VRAM + 8GB RAM floor; Windows needs virtual memory (page file) enabled, and the error page suggests 40GB+ free on each drive
  • Generation speed randomly tanks: some NVIDIA drivers after 532 are reported ~10× slower; the project suggests rolling back to driver 531
  • Model downloads blocked or slow: use --hf-mirror, or download weights manually into models/checkpoints
  • Port in use: --port 7861

Uninstall

rm -rf ~/Fooocus    # models, LoRAs and outputs live here — back up models/ and outputs/ first

AiBOX's Manage → Uninstall handles it.

More: GitHub README · troubleshoot · update log

Model / API key setup

AiBOX fills this in for you during the wizard — this section is for setting it up by hand.

OpenRouter

You need an API key from the provider.

~/SillyTavern/data/default-user/settings.json

Get a key ↗

Anthropic (Claude)

You need an API key from the provider.

~/SillyTavern/data/default-user/settings.json

Get a key ↗

OpenAI

You need an API key from the provider.

~/SillyTavern/data/default-user/settings.json

Get a key ↗

Google AI Studio

You need an API key from the provider.

~/SillyTavern/data/default-user/settings.json

Get a key ↗

Claude subscription

Sign in with your account instead — no key to copy.

~/SillyTavern/data/default-user/settings.json

Get a key ↗

Ollama (local)

Local model — no API key needed.

~/SillyTavern/data/default-user/settings.json

Common questions

Which platforms does Fooocus run on?

Runs on Windows / macOS — install it with AiBOX on any of them.

How hard is Fooocus to install?

Difficulty 4 out of 5. AiBOX does the download and setup; your part is clicking through the wizard and pasting a key if the app needs one.

Where does Fooocus keep my data?

Everything stays on your own machine: ~/Fooocus / ~/Fooocus. To uninstall, delete these — nothing is ever uploaded.

Something not working?

An outdated step, an error you do not understand, a typo — tell us and we will fix this page.

Report a problem →

Version & updates

Current version
v2.5.5
Registry data updated
2026-11-16

Version numbers are re-checked by our daily probe — when upstream ships a new release, this page follows.

Questions or problems?

A wrong step, a confusing error, a typo — open an issue and someone will look.