I spent about a week and a half building Fooocus Front, a Windows desktop app that installs, launches and drives Fooocus without a terminal, a console window or a setup guide.
Then someone told me about ComfyUI.
That annoying and sinking feeling is what this post is about, and if I'm being honest I still feel it as I'm writing this post. But having gone and checked properly, I'd say I probably have been worrying about the wrong thing.
What I built, briefly
Fooocus is excellent and deliberately minimal: one page, no sampler dropdowns, no CFG sliders. I'm a visual person and wanted something that feels like a place to make pictures, so I built one.
The interesting parts weren't the interface. They were a hidden launch (every Fooocus .bat ends in pause, which blocks forever with no console to press a key in, and Python block-buffers to a pipe so a quiet startup looks like it's frozen), a small Python bridge that runs inside Fooocus and puts jobs onto its own internal queue rather than scraping gradio, and an installer that sets up the right PyTorch build for whatever graphics card you own.
That last one taught me the most. The obvious way to check whether an Arc machine is already configured is to run python -c "import torch" and see if it works but it doesn't. Torch's Intel build needs a DLL search path the launcher sets up separately, so the probe fails on a perfectly healthy install. A check that reports a working setup as broken would trigger exactly the destructive reinstall it exists to prevent. It asks pip what's installed instead.
The release notes have the full feature list. This post is about what happened next.
The panic
"ComfyUI already exists" hit me about a day ago and it still stings. I built a UI. ComfyUI is a UI. Case closed.
Except when I actually checked:
ComfyUI is a node graph. I wanted something simpler than Fooocus's one-page form. ComfyUI goes in the opposite direction. Arguing I should have switched is arguing I should have solved a complexity problem by adopting more complexity.
Intel Arc is not a solved path there. I have an A770. Running ComfyUI on Arc means community install scripts or a third-party portable XPU build. Both are real options, both are maintained by people who aren't the ComfyUI team, and neither is "download and go."
SwarmUI is worse for my hardware. SwarmUI is the honest competitor, since it puts a friendly UI over a ComfyUI backend. But its maintainer's own answer in the Intel GPU discussion amounts to this: Swarm runs a stock ComfyUI as its backend, so install the IPEX prerequisites yourself and hope. Mine is verified on the card I own, and it generates from the moment the installer finishes.
So on the evidence, the week and a half I spent wasn't a duplicate. I'd like to say that settled it for me, but it hasn't quite yet.
The thing I should have been worried about
While writing this post I went and read the part of the Fooocus README I skipped like the fool I am.
Fooocus is in limited long-term support, bug fixes only, with no current plans to adopt newer model architectures. It's SDXL and it will stay SDXL. The README itself points people who want Flux towards WebUI Forge, ComfyUI or SwarmUI. The last release was v2.5.0, in 2024.
I built a careful installer, a repair system and a native interface for an engine that has stopped moving.
That's the actual critique, and nobody made it to me. The obvious objection, "ComfyUI exists," was wrong and easy to answer. The real one was sitting in the README I already opened.
There's a second one I found the same afternoon: my Intel Arc setup installs intel-extension-for-pytorch, which Intel retired at the end of March 2026 after upstreaming support into native PyTorch. My installer's headline feature is built on a package that's deprecated. Moving that path over to native PyTorch XPU is going into the next release.
Where that leaves it
Not in the bin. A few things survive the reckoning.
The deprecation is less dramatic than it sounds. The wheels still exist and still work, and Fooocus pins torch==2.1.0 anyway, which is exactly what my Arc branch installs. It isn't arbitrarily stale, it's matched to what Fooocus asks for. The migration also looks more tractable than I first assumed: RTX 50-series owners already run Fooocus on torch 2.7 and later to get Blackwell kernels, so the codebase tolerates a newer torch better than its pins suggest.
The bridge design matters more than I realised. Because it talks to Fooocus's internal queue rather than its web UI, retargeting it at a fork is a small change rather than a rewrite. Several forks exist, though check their commit history before trusting any of them; the one most often recommended hasn't shipped since 2024.
And SDXL being frozen isn't the same as SDXL being bad. It still makes good images, on hardware people already own.
The lesson, such as it is
"Check whether someone already built this" is not the useful question. Something has always already been built.
The useful question is whether the thing that exists solves your problem, and then, separately, whether the foundation you're building on is still moving. I got the first one right by accident and missed the second one completely.
If you're about to start something like this, don't be like me: read the project status section of the README before you write a line. It's usually two paragraphs, and it's usually the most important thing on the page.---
Fooocus Front v0.1.0-beta is out now. GPL-3.0, Windows only, unsigned installer, and Intel Arc is the only card it's been verified on. Every bit of the actual image generation is Fooocus, by lllyasviel and its contributors. They solved the hard part.