Fooocus: install the AI image generator free

#What is Fooocus?

Fooocus was released in August 2023 by lllyasviel, the author of ControlNet, under the GPL-3.0 license. The repository has surpassed 50,000 stars on GitHub. Its design principle: hide the technical details. Where other interfaces expose dozens of sliders, Fooocus automatically applies the settings that produce good results with SDXL and enriches your prompt behind the scenes using a GPT-2-based prompt-processing engine that runs offline. The README compares it to Midjourney, while noting that everything runs locally: no account, no subscription, and no image sent to a server.

Its status matters as much as its features. The README describes it as “limited long-term support (LTS) with bug fixes only,” with no plan to support new architectures. The latest release, 2.5.5, dates from August 12, 2024. Fooocus remains an excellent SDXL tool, not a platform that will keep up with tomorrow’s models.

• Understanding Stable Diffusion and its versions

#Requirements

The README provides a table of minimums by platform, along with speed comparisons from the project, most of them dating from 2023. Treat them as rough estimates, not current measurements.

Plan generously for disk space: the default model, juggernautXL_v8Rundiffusion, weighs 7.11 GB on Hugging Face, and the inpainting model downloaded for the first edit adds 1.28 GB. The README recommends at least 40 GB of free space on each disk if the “RuntimeError: CPUAllocator” error appears. With 4 GB of VRAM, Fooocus relies on Windows virtual memory, which is enabled in most cases: that is what makes execution possible—and slow.

#Install Fooocus

1. 01
   Download the archive

   From Fooocus's official GitHub page, follow the Windows download link: the .7z archive (Fooocus_win64_2-5-0.7z, about 2 GB) contains Python and all dependencies. The README warns that fake sites are circulating: download only from the lllyasviel/Fooocus repository.
2. 02
   Extract

   With 7-Zip, in a folder with a short path and no accented characters, such as C:\Fooocus. Avoid the Desktop and folders synced to the cloud.
3. 03
   Run run.bat

   On first launch, Fooocus automatically downloads its default model to Fooocus\models\checkpoints. If you already have these files, copy them there to save time. Allow ten minutes to half an hour, depending on your connection.
4. 04
   Open the interface

   The browser opens on its own at port 7865, defined in the launch code. If it doesn't, the address is displayed in the terminal's black window, which must remain open.

Two other launchers have been provided since version 2.1.60: run_realistic.bat loads realisticStockPhoto_v20, focused on photography, and run_anime.bat loads animaPencilXL_v500, focused on illustration. Since version 2.3.0, you can also change the preset in the browser. Each model is downloaded on first use. On Linux, the README offers installation through Anaconda or a venv with Python 3.10:

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
python entry_with_update.py

#Your first image

Describe the image in the bottom field, preferably in English, then click Generate. By default, Fooocus produces two images, according to its configuration code. Check Advanced to reveal the settings that matter.

Performance

Speed for everyday use, Quality for a final image, and three quick modes for iteration. The table below shows the number of steps for each mode.

Aspect Ratios

The image format. The code defines a list of SDXL formats, including 1152 × 896 by default and 1024 × 1024. Stick to these formats: SDXL was trained on specific dimensions and degrades outside them.

Styles

The default preset checks three styles: Fooocus V2, Fooocus Enhance, and Fooocus Sharp. Add only a few: stacking ten of them muddies the result.

Seed

Uncheck Random to fix the seed and reproduce the same image while changing only one prompt detail.

These modes are a tradeoff between time and quality: fewer steps means less refinement. Iterate in fast mode on the prompt, then rerun the best seed in Speed or Quality.

#The features that make the difference

Upscale or Variation

Upscale a successful image or request closely related variations.

Image Prompt

Provide reference images—four by default according to the configuration—to guide the style, composition, or a face, without training a model.

Inpaint or Outpaint

Repaint an area you mask with the mouse, or extend the image beyond its edges. The first use downloads Fooocus's dedicated inpainting model.

LoRA

Five SDXL LoRA slots are available in the Model tab. Place the files in the models/loras folder.

#Three prompting tips documented in the README

The README describes a little-known syntax, illustrated in the block below. Wildcards: a line such as __color__ flower picks a random color from the wildcards/color.txt file. Arrays: double brackets around a list produce one image per element, provided you set the number of images to 3 for three elements. Inline LoRAs: the lora tag, with the filename and a weight, applies a LoRA to the prompt, provided the file is in models/loras. These three syntaxes let you explore variants at scale without rewriting the prompt.

__color__ flower
[[red, green, blue]] flower
flower <lora:sunflowers:1.2>

#Models, configuration, and network access

After the first launch, Fooocus generates a config.txt file, where you can move model folders using keys such as path_checkpoints and path_loras. This is the right way to reuse models already downloaded for another interface without duplicating them. The README recommends thinking carefully before modifying this file and deleting it if it breaks: Fooocus then returns to its default values.

#Troubleshooting: RTX 50 and other errors

“no kernel image is available for execution on the device” (RTX 50)

The card is detected, but the bundled PyTorch does not know its architecture. The August 24, 2026 report on a portable RTX 5050 lists PyTorch 2.1.0 for CUDA 12.1 and supported architectures ending at sm_90. The fix documented by users is to install PyTorch for CUDA 12.8 in the python_embeded folder, but the project does not support this, and the report says Fooocus continued using the old version after the update. The most reliable option is to switch to ComfyUI.

“RuntimeError: CPUAllocator”

Windows virtual memory is disabled or insufficient. Re-enable it, and keep at least 40 GB free on each drive, the README recommends.

“MetadataIncompleteBuffer” or “PytorchStreamReader”

A model file is corrupted, often after an interrupted download. Download it again.

Generation far too slow

Check that the GPU is properly detected in the terminal. The README also notes that, in 2023, some NVIDIA drivers later than version 532 could be up to ten times slower than 531: test this only if nothing else explains the slowdown.

The RTX 50 pitfall applies to many other software packages: torch.cuda.is_available() can return True even though the first actual computation fails. Always test with a real computation, not just this indicator.

• Source: portable bug report RTX 5050 (August 24, 2026)
• Source: the RTX 5090 report (February 2025)
• Source: PyTorch 2.7 and Blackwell support

#Limits to know about in 2026

Fooocus remains limited to the SDXL architecture. Families released since then, such as Flux, Stable Diffusion 3.5, and Z-Image, are not supported, and the README itself points to WebUI Forge or ComfyUI/SwarmUI for recent models. The repository now receives only minor fixes: its last three commits, in January and September 2025, fix a path in the README, remove a link, and update a continuous integration tool. The README mentions forks worth trying, whose maintenance status should be checked case by case.

• ComfyUI: the logical next step when Fooocus is no longer enough
• AUTOMATIC1111: install Stable Diffusion WebUI
• Overview of local image-generation tools
• The AI Images Kit
• Fooocus’s official GitHub repository

#FAQ