Fooocus

Last reviewed Aug 26, 2026 · 999 words

Fooocus is the local image generator I hand to people who have never heard the word "sampler", and it is also feature-frozen: its author moved it into limited long-term support in 2024, bug fixes only, and it runs SDXL-family models exclusively. If you want to type a sentence and get a good picture within an hour of downloading, nothing self-hosted beats it. If you want to run this year's model families, you will outgrow it, and better to know that on day one than day thirty.

An hour from zero, most of it download time

On Windows the install is a single archive with an embedded Python and a run.bat; on Linux it is a Python 3.10 environment and one script:

git clone https://github.com/lllyasviel/Fooocus
cd Fooocus
python3.10 -m venv venv && source venv/bin/activate
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements_versions.txt
python entry_with_update.py --listen

First launch fetches the default SDXL checkpoint and supporting models, roughly 8 GB in total, then opens a Gradio interface on port 7865. entry_with_update.py pulls the latest code on every start; use launch.py once you have a version you like, so the UI does not change under you. Two pip commands and a clone, with no compose file, is why the catalogue rates it Easy despite the 8 GB RAM floor.

4 GB of VRAM survives, 8 GB is comfortable, and system RAM is the hidden floor

Fooocus does aggressive automatic model offloading, which is why the official minimum for Nvidia is 4 GB of VRAM where other SDXL front-ends want 8. The cost of that offloading is system RAM: the 8 GB in the catalogue is a working minimum, and 16 GB stops the swapping. Rough generation times for a 1024×1024 image at the default 30 steps, from my runs and reports I trust: about 20 seconds on an RTX 3060 12 GB, about 5 seconds on an RTX 4090, and a couple of minutes on a 4 GB GTX 1650. AMD works on Linux through ROCm and on Windows through DirectML, the latter slowly; Apple Silicon runs via MPS at minutes per image, fine for trying it, not for a workflow. The general method for matching a card to a model is in VRAM math, and for Fooocus the short version is that SDXL at fp16 is about 6.5 GB of weights, so anything above that stays resident and anything below streams.

The defaults are the product

What makes Fooocus different is that the author encoded a good SDXL workflow into defaults you never see. Prompt expansion runs your text through a small language model before generation, which is why "a cat" produces something a Midjourney user would recognise instead of a flat stock photo. Three styles are on by default (Fooocus V2, Enhance, Sharp), and the style picker offers around a hundred more. Performance presets set steps and scheduler for you: Speed is 30 steps, Quality is 60, and Extreme Speed, Lightning, and Hyper-SD drop to 4 to 8 steps for near-instant drafts at a quality cost. Launch with --preset realistic or --preset anime and the checkpoint, LoRAs, and styles swap to a tuned set for that look.

The Input Image tab is the other half. Upscale or Vary regenerates from an existing image; Image Prompt does style and composition reference without any training; and the inpaint and outpaint modes use a dedicated inpainting model that produces cleaner edits than the generic inpainting in older web UIs. Tick "Advanced" and seeds, guidance scale, sharpness, five LoRA slots, and the refiner appear, so the tool grows a little with you, but never into a node graph; the catalogue's "limited fine-grained control" con is accurate and intended.

On a server: --listen, auth.json, and a VPN

--listen binds to all interfaces, --port changes the default, and a file named auth.json in the Fooocus directory (copy auth-example.json) adds a username and password to the Gradio page:

[{"user": "studio", "pass": "a-long-random-passphrase"}]

That password is HTTP basic auth on an unencrypted Gradio server, which is adequate on a LAN and not adequate on the internet, so I put the port behind Tailscale rather than a public reverse proxy. Point path_checkpoints and path_loras in config.txt at the folder your other tools use and Fooocus will share one model library with ComfyUI or A1111 instead of duplicating 7 GB checkpoints.

The freeze, and what it means in 2026

The README says it plainly: Fooocus is in limited long-term support, and the author's attention has moved on. No Flux, no SD3.5, no newer families will arrive. SDXL fine-tunes on Civitai remain excellent for portraits, illustration, and product-style images, and for many people that is the whole requirement. If your reference point is what the newest models do with text in images, hands, or complex scenes, SDXL will feel a generation behind, because it is.

When you outgrow it, ComfyUI runs every model family and every technique at the price of a steep first day; InvokeAI is the polished middle ground with a canvas and a proper model manager. The side-by-side with the node graph is in ComfyUI vs Fooocus.

What I'd do

Install Fooocus first regardless of where you will end up, because two hours with it teaches you what a good SDXL image looks like before you have learned a single parameter name. Run it on an Nvidia card with 8 GB or more and 16 GB of system RAM, pin it with launch.py, gate it with auth.json over Tailscale, and share a model folder. Keep it as the quick-draft tool even after ComfyUI arrives; nothing else gets from idea to decent image in fewer clicks.