Fooocus

What is Fooocus?

Fooocus is a free, open-source AI drawing tool that can be used offline on local devices. It is based on Stable Diffusion XL; by simplifying the parameters and automating the optimization process, it allows users to focus on their prompts and the content of the images, without having to first master complex samplers, node diagrams, or workflow configurations.

The only official channel for releasing Fooocus is the GitHub repository maintained by lllyasviel. Various independent domain names on the Internet that carry the name Fooocus are not official sites; they may offer third-party hosting, subscription, or download services.

The download and installation procedures, version details, and security announcements should refer to the official repository.

Current maintenance status

Fooocus is currently in a phase of limited long-term support; the developers focus mainly on fixing errors and no longer work actively on developing new major features. The project is entirely based on SDXL, and there are no plans at present to migrate to or integrate new model architectures such as Flux.

This does not mean that the existing functions cannot be used. Users who need stable local SDXL generation, a simple interface, and a well-developed image editing process can still choose Fooocus.

Users who wish to use Flux, the latest video models, or automation for complex nodes would be better off evaluating ComfyUI, WebUI Forge, or other tools that are still under development.

Text-to-image generation and automatic prompt expansion

SDXL images can be generated once the user enters a natural language prompt. Fooocus takes care of aspects such as sampling, resolution, model loading, and quality parameters automatically, and it also offers common settings like style, image size, number of images, quality, and negative prompts.

Fooocus V2 utilizes a local GPT-2 model to expand on the given prompts, adding visual details to brief descriptions. This process can be carried out offline, but the automatic expansion may not always align with the characteristics of each brand or character.

When a stable reproduction is required, it is necessary to save the complete prompt, seed, model, LoRA, and version information.

Style presets and prompt control

The built-in style system allows for quick application of various visual styles such as photography, cinema, illustration, and anime. It also supports multiple prompt lines, negative prompts, and weight syntax; for example, increasing the weight of a certain concept can make the features of that element more prominent.

Predefined options are suitable for quick exploration, but the combination of multiple styles can lead to conflicts. Character consistency, accurate text, details of the hands, and complex spatial relationships still require repeated creation, partial adjustments, or the use of reference images.

Comparison between the official version and preset versions

Fooocus does not offer any official cloud subscriptions, point packages, or enterprise plans. The cost of the local software is zero, but expenses are incurred for graphics cards, electricity, storage, and third-party GPU hosting.

The fees charged by any online websites with the same name do not represent Fooocus’ official pricing.

Image variants and zooming

The Upscale or Variation option in the Input Image allows for creating slight or significant variations of an existing image; it is also possible to enlarge the image by 1.5 times or 2 times. Slight variations keep the image closer to its original form, while significant variations reinterpret various features and details of the image.

AI upscaling is not just traditional interpolation; it may involve redrawing textures, facial features, and text. Documents, details of products, logos, and images that require pixel-level accuracy must be checked individually after the output is generated.

Local redrawing and image expansion

Inpaint allows users to select the areas that need to be modified, and then use prompts to replace objects, repair facial features, adjust clothing, or remove defects. Outpaint enables the expansion of the canvas in all directions – up, down, left, and right – so that the AI can fill in the content outside the frame.

Fooocus employs its own algorithm for local redrawing as well as its own control model. The first time that these functions are used, an additional 1.28 GB of data containing the local redrawing patches will be downloaded; therefore, it is necessary to have sufficient network and disk space available.

Images with complex boundaries, strict perspective requirements, or continuous structures across different regions may require multiple masking steps and post-production compositing.

Reference image for the image prompt

Image Prompt can be used to guide the composition, character features, style, or specific elements of an image by providing reference pictures, and it also offers control capabilities related to ControlNet. The higher the intensity of the reference, the more similar the resulting image will be to the original one; however, this may also limit creative variations or amplify any flaws present in the original image.

When using real-person photos, brand materials, and copyrighted characters, users must ensure that they have the necessary permissions regarding portrait rights, trademarks, and model licensing; the generation tool does not obtain such commercial approvals automatically.

FaceSwap face swapping

Fooocus integrates FaceSwap, which is based on InsightFace technology, allowing the features of a reference face to be applied to the generated image. It is suitable for creating creative avatars and character sketches; however, side views, obstructions, unusual lighting conditions, and age differences can reduce the level of similarity.

It is not allowed to create deceptive identity content without consent. When real people, advertisements, media releases, or commercial purposes are involved, explicit authorization must be obtained and the necessary information regarding AI processing must be indicated.

Describe – reverse inference of image description

The Describe function can analyze an image and generate a natural-language description or prompts suitable for drawing, which helps in understanding the elements of the image and enables quick reconstruction of similar versions or further creation of variations. The output obtained is a summary of the image by the model; it does not reproduce the original creator’s full prompts, nor the model parameters.

Enhance intelligent enhancement

The Enhance function makes use of object detection and segmentation techniques to dynamically locate specific areas, allowing for the refinement or repair of those designated objects. By combining the capabilities of GroundingDINO and SAM, this process is suitable for improving the quality of faces, hands, clothing, or other individual objects.

Automated detection may result in items being missed, incorrectly selected, or having their original characteristics altered. Before processing in batches, it is necessary to use representative images to test the range of the mask, as well as the enhancement prompts and their intensity.

LoRA, Embeddings, and Models

Fooocus supports SDXL models, LoRA, and Embeddings; it also allows the configuration of multiple model directories as well as the use of SDXL-compatible resources from platforms such as Civitai. LoRA is useful for adding details to characters, clothing, styles, and object concepts, while Embeddings are often used for controlling prompts or negative aspects.

Each model and LoRA comes with its own separate license; an open-source software license does not automatically grant commercial rights to use the model. When downloading third-party models, it is also important to pay attention to malicious files, trigger words, the version of the base model, and the recommended weights.

Wildcards and batch changes

Wildcards can randomly replace elements from a vocabulary, while array processing and inline LoRA enable users to create combinations of clothing, colors, lenses, and characters. It is suitable for exploring ideas and generating large quantities of material, but the number of combinations can increase rapidly; it is therefore necessary to control the scale of tasks and keep track of the successful parameters.

Windows installation guide

1. Go to the release page via the only official GitHub repository and download the Windows archive.
2. Unzip the file completely into a directory with a short path and sufficient storage space.
3. Run run.bat to activate the general preset; use the corresponding startup file when anime or realistic default models are required.
4. Upon first launch, the SDXL model will be downloaded automatically; once this is complete, the local interface can be opened in the browser.
5. Before generation, verify the model, format, quality, and output directory; first test with a small number of images.

The model files downloaded for the first time are large in size, and antivirus software, proxies, disk permissions, or network interruptions can all cause failures when trying to start them. Do not download repackaged programs from unknown websites with similar names.

Linux, macOS, and Docker

Linux allows for the use of Python 3.10 as specified in the official guidelines to install dependencies and start the application; it also supports deployment via Docker. Apple Silicon Macs can run PyTorch MPS, but this is not covered by the official guidelines, and the performance is usually lower than that of standalone NVIDIA GPUs.

For AMD GPUs, one can try DirectML on Windows and ROCm on Linux; both are approaches whose compatibility and performance are less certain. Whether they will work depends on the GPU, drivers, operating system, and version of PyTorch.

Refer to hardware requirements.

Actual requirements are also influenced by resolution, batch size, magnification, ControlNet, and model size. The system should use drivers that are compatible with the current PyTorch version; it is not advisable to rely on drivers from older versions.

Model download and disk space

The generic, realistic, and anime presets will download different SDXL models, while functions such as local redrawing, zooming, object detection, and segmentation will also require the download of additional files as needed. To use them fully, several dozen GB of storage space should be reserved, and it is advisable to avoid storing the model directories on system drives with insufficient capacity.

LAN access and security

The startup parameters allow the interface to monitor LAN addresses, or the sharing functionality of Gradio can be utilized. In both cases, authentication is not enabled by default, and such interfaces should not be exposed to the public internet.

When remote access is required, basic authentication via auth.json should be enabled, along with the use of firewalls, reverse proxies, HTTPS, and access controls.

Uploading files, models, and extensions can all pose risks. In multi-user environments, it is necessary to isolate the output directories, restrict the sources from which files are uploaded, keep dependencies up to date, and avoid running tasks with administrator privileges.

Open-source license

The Fooocus code is licensed under the GPL-3.0 license. When modifying and distributing the program, it is necessary to comply with the requirements of the GPL regarding the source code and the license.

Ordinary images generated by software do not automatically become GPL-licensed just by using Fooocus; the rights to such images still depend on the input materials, the license of the model, the LoRA license, and local laws.

Fooocus usage guide

Create reusable professional workflows

1. Create a list of brand colors, fonts, layout elements, and elements that are prohibited.
2. Test separately the current maintenance status, text-to-image generation, automatic prompt expansion, as well as style presets and prompt control.
3. Use the same set of representative samples to compare quality, speed, and cost;
4. Complex edges, text, and images of key products should be handed over for manual refinement.
5. Standardize naming, dimensions, and review status;
6. Batch processing and release are carried out after random inspections;

Which users is it suitable for?

• Beginners in AI art generation who wish to run SDXL for free on their local devices;
• Illustrators and content creators who do not want to create complex node diagrams;
• Users who need text-to-image generation, image enlargement, partial redrawing, and image scaling;
• Individual users who place importance on offline processing and the privacy of their data;
• Users who use NVIDIA graphics cards with moderate to low video memory for lightweight generation;
• Teams that need to create anime, realistic, and general concept art quickly.

Main advantages

• It is free, open-source, and can run offline;
• The interface and parameters are easier to use compared to tools for complex nodes.
• Image generation, variations, enlargement, redrawing, and image expansion are all available in one interface;
• It offers automatic prompt expansion, style presets, and various controls for reference images.
• Supports LoRA, Embedding, FaceSwap, and image descriptions;
• It provides the minimum viable configuration for NVIDIA devices with 4GB of video memory.

Restrictions and Precautions

• At the moment, the project only receives limited long-term maintenance; it is focused primarily on SDXL, and no new architectures such as Flux will be integrated into it.
• Compared to ComfyUI, Fooocus is simpler, but it offers fewer options for custom nodes, automation, and underlying control.
• The first download of the file is large in size, and the performance of AMD, Apple Silicon, and CPUs is limited.
• The generated results may exhibit issues with consistency in hands, text, structure, and characters;
• The user is responsible for verifying the copyright, security, and commercial licensing aspects of third-party models, LoRA, reference images, and real-person materials.

Frequently Asked Questions

Is Fooocus free?

It’s free. The official local software does not require any subscriptions, bonus packs, or paid cloud services; users only need to bear the costs related to their hardware, storage, and electricity.

What is the official website of Fooocus?

The only official source is the GitHub repository maintained by lllyasviel. Several independent domains with the same name have been explicitly stated by the authorities as unrelated websites.

Does Fooocus support Flux?

The official version does not support this. The project is based on SDXL, and there are no plans at the moment to migrate to or integrate a new model architecture.

Can it run with 4GB of video memory?

The minimum requirements specified by NVIDIA are 4GB of video memory, 8GB of RAM, and the use of swap space; however, high resolutions, multiple images, and complex editing tasks significantly increase the resource needs.

Can Fooocus perform local redrawing and image zooming?

Yes, it supports partial redrawing of masks and expanding the canvas in four directions; an additional file containing the models for partial redrawing will be downloaded the first time it is used.

Is Fooocus open source?

It is open source, with the code licensed under GPL-3.0; the models, LoRA components, and input materials are each subject to their own separate licenses.