Fooocus: Offline SDXL Image Generator
/Fooocus
Fooocus is a local Stable Diffusion XL image generator that trades ComfyUI-style node graphs for three buttons and a text box. Reach for it if you want SDXL-quality images without learning samplers, CFG scales, or a node editor. Skip it if you need Flux or a newer architecture — the project is in bug-fix-only mode with no plans to migrate.
What is Fooocus?
Fooocus is an offline, open-source image generator built on Stable Diffusion XL, packaged as a single downloadable app with a Gradio web interface. It borrows Midjourney's approach of typing a prompt and getting a finished image, but runs entirely on your own GPU instead of a cloud API. Where Automatic1111 and ComfyUI expose every sampler and node, Fooocus hides that machinery behind presets and an internal GPT-2-based prompt-processing engine.
Fooocus features at a glance
- ✓An offline GPT-2-based prompt processing engine that reworks your prompt internally so results stay coherent whether you type "house in garden" or a page-long description.
- ✓Its own inpaint algorithm and a dedicated inpaint model (a separate ~1.28GB checkpoint) instead of the standard SDXL inpaint pipeline.
- ✓Its own image-prompt algorithm for image-to-image guidance, which the README describes as more faithful than standard SDXL IP-Adapters or Revisions.
- ✓Three ready-made presets — Default, Anime, Realistic — each loading a different base checkpoint (juggernautXL_v8Rundiffusion, animaPencilXL_v500, realisticStockPhoto_v20) via a separate launcher script or a --preset flag.
- ✓Prompt weighting syntax borrowed from Automatic1111, like "(word:1.5)", plus embedding syntax, so prompts copied from Civitai work unmodified.
- ✓FaceSwap via InsightFace, reached through Input Image -> Image Prompt -> Advanced.
- ✓Vary, upscale, inpaint, and outpaint all chained from one "Input Image" panel instead of separate tools.
Installing Fooocus
Windows: download the packaged .7z release, unzip it, and run run.bat — Fooocus downloads its default SDXL checkpoint on first launch. Linux: clone the repo and either build a conda environment from environment.yaml or a venv against Python 3.10, then run entry_with_update.py; add --listen to expose it on the network. Mac (Apple M1/M2): install PyTorch nightly with MPS support, clone the repo, build the conda environment, and run entry_with_update.py — the README calls this path unofficial and notes Apple silicon has no dedicated GPU, so generation runs much slower. AMD GPUs are beta. Windows uses a DirectML build of run.bat; Linux reinstalls torch against the ROCm wheel — the README says AMD support is "not intensively tested." Colab has an official notebook that launches with --share --always-high-vram, with the refiner disabled by default because free-tier resources are limited. Docker install is covered in a separate docker.md file in the repo.
Strengths
- ✓Install-to-first-image is genuinely short: unzip, run run.bat, and the app fetches its own checkpoint.
- ✓Vary, upscale, inpaint, outpaint, and image-prompt all live in one "Input Image" panel instead of separate tabs or nodes.
- ✓Prompt weight and embedding syntax matches Automatic1111, so prompts and weights copied from Civitai work without rewriting.
- ✓Free and fully offline once the models are downloaded, with no account or API key involved.
- ✓Three presets swap the whole model and config stack — general, anime, realistic — with one flag or launcher script instead of manual checkpoint juggling.
Limitations and known constraints
- △The README indicates the project currently sits in "Limited Long-Term Support (LTS)" — there are no current plans to add features or migrate to newer architectures like Flux, and bug fixes only are being addressed.
- △AMD support is beta. The README calls it "not intensively tested" for both the DirectML build on Windows and the ROCm build on Linux.
- △Mac support is unofficial, community-written instructions, and Apple silicon runs roughly 9x slower than an Nvidia RTX 3XXX card per the README's own benchmark table.
- △Minimum viable hardware still needs a 4GB Nvidia GPU plus 8GB system RAM with swap enabled; CPU-only mode works but the README puts it at about 17x slower than an RTX 3XXX.
- △--listen and --share expose the UI unauthenticated by default — you have to hand-create an auth.json with user/pass entries yourself to lock it down.
Alternatives to Fooocus
Who should use Fooocus
Fooocus fits a hobbyist or solo creator who wants SDXL-quality images on their own GPU without learning ComfyUI's node graph or Automatic1111's parameter panels — install, type a prompt, done. It also suits anyone who wants Midjourney-style output but needs it offline and free. It's the wrong pick if you need Flux or another newer architecture, want node-level pipeline control, or need a project still receiving new features rather than bug fixes.
Frequently asked questions
Fooocus needs an Nvidia GPU with at least 4GB of VRAM and 8GB of system RAM, with Windows' virtual swap enabled to cover the gap. Nvidia RTX cards and GTX cards from the 9XX/10XX series onward are supported; anything older than a GTX 900-series card is not.
Yes — Fooocus is free, open-source software under the GPL-3.0 license, and it runs entirely offline once its default SDXL checkpoint has been downloaded. No account, API key, or internet connection is needed to generate images afterward.
Fooocus supports AMD GPUs through a beta DirectML build on Windows or a ROCm build on Linux, though the README calls AMD support "not intensively tested." It also runs on Apple M1/M2 Macs via PyTorch's MPS backend through an unofficial, community-documented path, and falls back to CPU-only mode if no GPU is available.
The three Fooocus presets each load a different base checkpoint: Default uses juggernautXL_v8Rundiffusion for general images, Realistic uses realisticStockPhoto_v20, and Anime uses animaPencilXL_v500. You switch between them with run.bat, run_anime.bat, or run_realistic.bat on Windows, or the --preset flag on Linux and Mac.
Fooocus's production readiness isn't clearly addressed beyond its GPL-3.0 license — the README doesn't discuss commercial terms. The project is also now in "Limited Long-Term Support (LTS) with bug fixes only," with no new features planned, which matters if a production use case needs ongoing development.
Fooocus trades the node graph of ComfyUI and the parameter-heavy tabs of Automatic1111's stable-diffusion-webui for three buttons and a text box, hiding sampler and prompt tuning behind presets. The tradeoff is control: Fooocus's own README recommends ComfyUI or WebUI Forge instead once you need newer architectures like Flux, since Fooocus itself won't migrate.
The problem it solves
Getting good SDXL output normally means hand-tuning samplers, CFG scale, and prompt weights across a node graph or a dozen UI tabs — and a prompt copied from Civitai often breaks because different UIs use different reweighting syntax. Fooocus's README frames its whole design around collapsing that setup into three clicks and an offline prompt-processing engine that keeps results stable whether the prompt is two words or two hundred.
Best use cases
- •Generating SDXL images without learning sampler, CFG, or node-graph settings.
- •Iterating on a single image via vary (subtle/strong), upscale (1.5x/2x), inpaint, or outpaint from the same panel.
- •Running anime- or realistic-photo-style generation through the dedicated presets instead of manually swapping checkpoints.
- •Face-swapping a generated portrait using InsightFace.
- •Testing SDXL checkpoints and prompts pulled from Civitai without rewriting prompt weight syntax.
Related repositories
Still deciding about Fooocus?
One click hands the question to an AI along with this page — see what it says about Fooocus.