Fooocus Front v0.1.0-beta

First beta of Fooocus Front, a Windows desktop app that installs, launches and drives
Fooocus.

All of the image generation here is Fooocus, by lllyasviel and
its contributors. They built the hard part and built it well; this project only adds a different
way to sit in front of it. Fooocus's own interface is deliberately minimal and, for a lot of
people, exactly right — so it stays one click away inside the app, and always will.

This exists because I am a visual person and wanted somewhere that feels like a place to make
pictures: a canvas that fills the window, previews building as they render, models as a browsable
library, outputs as a gallery. That is a difference in taste, not quality.

The second reason is installation. Getting any Stable Diffusion tool running means matching a
specific PyTorch build to whatever graphics card you own, and that is genuinely awkward for
everyone shipping in this space rather than a flaw in Fooocus. It is no trouble at all if you are
technical, and a wall if you are not. This app does that part for you.

Read this before installing

• Windows only. There is no Mac or Linux build.
• The installer is not code-signed. SmartScreen will warn you. Choose More info -> Run
  anyway. If that is not acceptable to you, do not install it.
• AMD graphics are untested. Setup follows the official Fooocus instructions for AMD but has
  never been run on an actual AMD card, because no AMD machine was available. The app says so in
  the interface. NVIDIA and Intel Arc are handled; Arc is what this was developed and verified on.
• You need roughly 15 GB free for Fooocus plus the essential models.

What it does

Setup detects your graphics card, downloads and extracts the official Fooocus standalone
package, configures the right stack for your hardware, and offers the seven essential models with
a real progress bar. If you already have Fooocus, point it at your existing folder instead.

Studio is the native interface: prompt, aspect ratio, performance, seed, 279 styles with
spelling-tolerant search, and LoRAs with weight sliders. Inpaint and outpaint have a proper mask
editor with free zoom and undo. Upscale and Vary covers all five methods. Image Prompt gives four
slots across ImagePrompt, PyraCanny, CPDS and FaceSwap.

Write prompts in your own language. Pick from ninety-eight languages and type prompts in
yours — they are translated into English before generating, and the English is shown alongside so
you can see exactly what was sent rather than having your words silently rewritten. This matters
because Stable Diffusion understands English far better than any other language, so translating
prompts does more for your images than translating the buttons would.

Choosing a language costs nothing. Its model — about 300 MB — downloads when you ask for it, once,
and can be cancelled or removed later. Nothing is bundled into the installer, and an English user
downloads nothing at all. Languages already downloaded are ticked in the list.

Fifty-nine languages have a model trained on that exact pair. The rest use the closest
language-family model, and four fall back to a general one. The app tells you which you are
getting, because a weaker model drops words rather than failing outright.

Models browses what you have installed and searches Civitai, filtered to SDXL, with content
controls and detection of models you already own. Downloads are resumable and survive a restart.

About the Civitai API key

Browsing Civitai works without an account. Downloading from it needs a free API key from your
Civitai profile, which is their rule rather than mine — they have required it since 2024.

You enter it once. It is verified against Civitai before being saved, then kept in Windows
Credential Manager rather than in a settings file. It survives restarts, and reinstalling the
app does not clear it — the credential store is not part of what an uninstall removes, so a fresh
install on the same machine finds Civitai already working. A new machine, or a different Windows
account on the same one, needs it entered once more; Windows encrypts the entry against the
account that saved it, which is the point of keeping it there.

Since an unsigned app asking for an API key is a reasonable thing to be wary of, what to check:
the key is only ever sent to civitai.com and the Cloudflare bucket their downloads redirect to,
and a Hugging Face download is explicitly given no token rather than trusted not to need one.
Every request is made from Rust, so the key is never handed to the webview — the interface can ask
whether a key exists and gets back yes or no, never the value.

The complete list of hosts this app can contact is api.github.com, github.com,
huggingface.co, civitai.com and pytorch-extension.intel.com. There is no telemetry or
reporting of any kind, and the Python bridge running inside Fooocus makes no outbound connections
at all. All of that is in the source, and the README says where to look.

Anything the native interface does not cover is still there behind the Fooocus tab.

Known limitations

• The app's own interface is English only. Prompts are translated; the buttons are not.
• The from-scratch installer has been run end to end on one clean machine besides the development
  one. That single run found three real bugs, all fixed, so expect more on hardware nobody has
  tried yet.

Bug reports are welcome, particularly from anyone installing on hardware or a Windows
configuration different from the two machines this has seen.

---

Licensed GPL-3.0, matching Fooocus
itself. Translation uses the Helsinki-NLP OPUS-MT models (Apache-2.0). This is an independent
front end and is not affiliated with the Fooocus project. Please send bugs here rather than to the
Fooocus maintainers — anything wrong with this app is mine, not theirs.