lllyasviel/Fooocus

What it is

Fooocus is a Gradio-based, offline image generator built entirely on Stable Diffusion XL. Its goal is to make text-to-image generation behave more like hosted services such as Midjourney: users write prompts and look at images rather than tuning samplers, CFG, and refiner settings. A built-in GPT-2 prompt expansion engine and tuned sampling defaults handle much of that work, so both a three-word prompt and a long prompt can produce usable output. It also bundles image prompt, inpaint/outpaint, upscale/variation, face swap, and style presets behind one local web UI, aimed at people who want local SDXL output without a node graph or a dense parameter panel.

The project describes itself as being in limited long-term support, with bug fixes only and no current plan to adopt newer model architectures. Users who need Flux or newer models are pointed to WebUI Forge, ComfyUI/SwarmUI, or community forks instead. The README also warns that several similar-looking domains are fake and that the GitHub repository is the only official source.

Quick start

The easiest path is the Windows package: download the v2.5.0 Windows archive, extract it, and run run.bat. On first launch it automatically downloads the default checkpoint (juggernautXL for the general preset) and later the inpaint patch when you first use inpainting; models land under Fooocus\models\checkpoints. The browser UI opens and you can generate from a text prompt. run_anime.bat and run_realistic.bat select the anime and realistic presets, each with their own automatic model download.

On Linux with Anaconda:

git clone https://github.com/lllyasviel/Fooocus.git
cd Fooocus
conda env create -f environment.yaml
conda activate fooocus
pip install -r requirements_versions.txt
python entry_with_update.py

Add --listen to expose the UI on the network, or use --preset anime / --preset realistic. A Python 3.10 venv or system Python path is also documented. AMD GPUs require replacing torch with the ROCm build on Linux or with torch-directml on Windows; Mac M1/M2 runs through PyTorch MPS. A Colab notebook is available for free-tier testing, with the refiner disabled by default there. Docker instructions live in docker.md.

Hardware requirements

The minimum is a 4GB Nvidia GPU and 8GB system RAM on Windows/Linux, with system swap enabled. AMD needs 8GB VRAM (DirectML or ROCm, both marked beta); Apple silicon uses shared memory and is roughly 9x slower than an RTX 3XXX according to the project's table. CPU-only operation needs 32GB RAM and is about 17x slower. The project explicitly does not support smaller models to lower these requirements. Default model downloads are several GB, and first launch can take a long time on slow connections.

Remote access and configuration

--listen and --share expose the UI without authentication by default. To add HTTP basic auth, create auth.json in the main directory with a list of { "user", "pass" } objects, following auth-example.json. After the first launch, Fooocus\config.txt holds model paths and default parameters; deleting it restores defaults. Prompt syntax supports wildcards like __color__, arrays like [[red, green, blue]] for batch variants, and inline LoRAs like <lora:sunflowers:1.2> when the file is in models/loras. Several command-line flags control VRAM strategy, attention backend, previews, language, and preset behavior.

HotRepo's take

The repository has built a large following (53,825 stars, 8,625 forks) since August 2023, and the project's own limits matter more than the count: it is frozen on SDXL and maintained for bug fixes. The 2023 Hacker News thread drew 186 points and 54 comments, with commenters reacting to the single-sampler refiner swap and the broader idea of a simpler local SDXL UI. One frontend developer in that thread found the sampling description impenetrable, which is a fair reminder that Fooocus hides complexity in defaults rather than in documentation. Its clearest value is for users who want offline SDXL output with minimal setup on modest Nvidia hardware; anyone who needs current model families should start with the alternatives the README names.