Fooocus

What it is

Fooocus is image generating software you run on your own computer. The README says it is offline, open source and free, and that you only focus on the prompt and the pictures, as with online generators such as Midjourney. Manual tweaking is not needed. It is built on Stable Diffusion XL and Gradio, so you use it in your browser. An offline GPT-2 based prompt engine expands a short prompt such as “house in garden”, so the result still looks good. From the same page you can upscale a picture, make variations, inpaint or outpaint, use an image as a prompt, or ask for a description of an image. The README says the project is in limited long-term support: bug fixes only, and no current plan to move to newer model architectures.

1. You type a prompt and press Generate
2. With the Fooocus V2 style, a GPT-2 model expands your prompt
3. Stable Diffusion XL draws the images on your computer
4. You upscale, vary or inpaint the one you like
5. Images are saved in the outputs folder

Who it suits

People who want pictures from text on their own computer, offline and free. It suits you if you know Midjourney or Leonardo, because the README maps their features to Fooocus’s. It also suits anyone with an Nvidia card who wants a simple browser interface over Stable Diffusion XL.

Skip it if you need newer models such as Flux. The README sends those users to WebUI Forge and ComfyUI. Skip it too if you have no GPU. The README’s table puts a CPU-only run at about 17 times slower than an Nvidia RTX 3XXX and asks for 32 GB of memory.

Where to use it

The README, and the Docker page it links, have a route for each of these.

• Windows with an Nvidia GPU: a downloadable archive and run.bat
• Windows with an AMD GPU: DirectML, through an edited run.bat, in beta
• Linux: Anaconda, a Python venv or the system Python
• Linux with an AMD GPU: ROCm, in beta
• Mac with Apple silicon: an unofficial guide, and much slower
• Google Colab: the official notebook, with a last line you can edit
• Docker, Docker Compose and Podman: an image for Nvidia GPUs

How to set it up

Get it saves a zip of Minda’s setup notes in your downloads folder, with a page for each platform above, one for models and config, and one for troubleshooting. On Windows with an Nvidia card, the README says to download the archive, uncompress it and run run.bat. On Windows with an AMD card you edit run.bat in the same archive. On Linux and Mac you clone the repository and install its Python packages. Colab runs the official notebook, and Docker or Podman run the ghcr.io/lllyasviel/fooocus image.

1. Check your GPU first. Compare it with the README’s table of minimal requirements before you download anything.
2. Get the code. On Windows, download the archive from the repository. On Linux and Mac, clone the repository.
3. Install it. On Windows, uncompress the archive and run run.bat. On Linux and Mac, install the packages, in a conda or venv environment unless you use the system Python.
4. Launch it and wait. The first launch downloads the default models for you.
5. Type a prompt on its page. Open the page in your browser first.

git clone https://github.com/lllyasviel/Fooocus.git
cd Fooocus
python3 -m venv fooocus_env
source fooocus_env/bin/activate
pip install -r requirements_versions.txt

That is the README’s venv route for Linux. It launches with these two lines.

source fooocus_env/bin/activate
python entry_with_update.py

How to use it

Type a prompt and press Generate. Most other controls sit in two places. Input Image holds what you do with a picture: upscale it, vary it, inpaint or outpaint it, use it as an image prompt or ask for a description. Advanced holds the style, quality, image number, negative prompt, aspect ratios and guidance. The README maps Leonardo’s Prompt Magic to the Fooocus V2 style under Advanced.

Example

You want one flower in three colors. Type the line below in the prompt box and raise the image number to 3. Fooocus makes one image for red, one for green and one for blue.

[[red, green, blue]] flower

A wildcard works the other way round. The prompt __color__ flower picks one random color from the file wildcards/color.txt, and the pick follows the seed. A LoRA goes in the prompt too: flower <lora:sunflowers:1.2> applies it when its file is in the models/loras folder.

For newcomers and experts

If you are new. Fooocus is a program, not a website. A model is the file that holds what the drawing system learned, and your graphics card does the work. Fooocus fetches the models it needs when you first start it. The README warns that fake Fooocus sites exist, so get it from the repository. On Windows, run_realistic.bat and run_anime.bat load different models than run.bat. Start with run.bat, write a plain sentence and press Generate. If the console shows “RuntimeError: CPUAllocator”, a crash such as Segmentation Fault or Killed, or closes in a flash, the troubleshooting page blames too little system swap. A “CUDA Out Of Memory” on an Nvidia card with 8 GB, or 4 GB with Float16 support, is a bug to report. On 6 GB it says very likely a bug.

For experts. The README lists the sampling work as tricks, and says they are based on SDXL and are not very up to date with the latest models. They include a refiner swap inside a single k-sampler, negative ADM guidance, a tuned form of self-attention guidance, and CFG scale with TSNR correction above 10. Prompt weights use A1111’s reweighting, which the README says suits prompts copied from Civitai. After the first launch, config.txt sets the defaults, and the usage line lists many flags, from --always-high-vram to --preset and --language. The lines below are an excerpt of the README’s example of an edited config.txt.

{
    "default_model": "realisticStockPhoto_v10.safetensors",
    "default_cfg_scale": 3.0,
    "default_sampler": "dpmpp_2m",
    "default_scheduler": "karras"
}

Good to know

Fooocus runs on your machine, and a few paths reach out.

• Status: the README says it is in limited long-term support, with bug fixes only. For newer models it names WebUI Forge, ComfyUI and SwarmUI
• Official source: the README says its GitHub page is the only official source. It lists fooocus.com, fooocus.net, fooocus.co, fooocus.ai, fooocus.org, fooocus.pro and fooocus.one as fake sites with no tie to the maker
• Network: --listen and --share expose the interface, and both are unauthenticated unless you add an auth.json file with a user and a pass for each login
• Downloads: the first launch downloads models, and the README names a Hugging Face address for the inpaint model. It does not cover the licenses of the models you download
• Flags: the README’s flag list includes --disable-analytics, --disable-image-log and --disable-metadata, and does not describe them
• Hardware: Nvidia is the main route. AMD support is in beta, Mac is unofficial, and the troubleshooting page blames too little system swap for most crashes
• Removal: the README gives no steps. On Windows you delete the folder you extracted
• License: GPL-3.0. The LICENSE file is the standard GNU General Public License version 3 text, with no copyright line for the maker