How to Use Fooocus for Uncensored Local Generation

The Quick Answer

Install Fooocus from the official GitHub repository, use a compatible SDXL checkpoint, and leave Black Out NSFW disabled. That's it.

Old guides tell you to edit modules/safety_checker.py or change a check_nsfw() function. Don't do that. Current Fooocus config sets default_black_out_nsfw to False, and the censor path only runs when the global or task setting is enabled.

The setting controls an output-image blackout. It isn't a documented cloud-style prompt rejection system or account policy.

What “Uncensored” Means Here

Fooocus runs the generation pipeline locally after the required software and models are on your machine. Your workflow doesn't depend on a hosted image service approving each prompt.

That doesn't guarantee every checkpoint will produce every request. A model's training, prompt understanding, and license still matter.

Use content you have the legal right to create. Local software doesn't remove consent, copyright, or model-license requirements.

What You Need

• Official Fooocus: Download only from github.com/lllyasviel/Fooocus. The project warns that lookalike Fooocus websites are fake.
• An SDXL checkpoint: The Windows presets download default models automatically, or you can add a compatible community SDXL checkpoint.
• Windows hardware: The official basic path lists an Nvidia GPU with 4 GB VRAM and 8 GB system RAM.
• Free disk and swap: Model files are large, and Fooocus troubleshooting repeatedly points to swap and free-space shortages.
• No extension: Basic local generation and the supported blackout setting don't require a plugin.

Fooocus is an SDXL application. Copying a Flux or SD3 file into its checkpoint folder won't add support for that architecture.

Step 1: Download the Real Fooocus

Go to the official lllyasviel/Fooocus GitHub repository and follow its release link. Don't download installers from fooocus.com, fooocus.net, fooocus.ai, or another lookalike domain.

On Windows:

1. Download the official Fooocus archive.
2. Extract it with 7-Zip.
3. Run run.bat for the default workflow.
4. Use run_anime.bat or run_realistic.bat if you want the official preset and model combinations.
5. Wait for the first-run model downloads to finish.

Linux and macOS work through the official manual installation paths. Windows with Nvidia hardware remains the simplest supported route.

Step 2: Keep Black Out NSFW Off

Current official source leaves the NSFW blackout disabled by default. You shouldn't need to change anything on a clean install.

If you changed settings before:

1. Open Fooocus and expand the advanced or developer controls.
2. Find Black Out NSFW.
3. Make sure it is disabled.
4. Check any persistent config for default_black_out_nsfw.
5. Set that value to false through configuration, then restart Fooocus.

Don't patch the censor module. A source edit can be overwritten by updates and makes later troubleshooting harder.

Step 3: Add the Right SDXL Checkpoint

The default Fooocus presets can download their own SDXL models. For a community checkpoint, confirm that it is built for SDXL, not SD 1.5, Flux, or SD3.

Place compatible checkpoint files in Fooocus's documented checkpoints directory, then refresh the model list or restart the application. Match any LoRA to the same base architecture.

Review the checkpoint's model card before use:

• Base architecture: It should state SDXL compatibility.
• License: Check personal, commercial, and redistribution terms.
• Trigger words: Some fine-tunes need specific phrases.
• Recommended settings: Resolution, sampler, and prompt style can change the result.

Step 4: Generate and Keep the Workflow Local

Choose your checkpoint, enter a prompt, and generate normally. Fooocus handles its prompt expansion and SDXL defaults without a node graph.

Useful built-in tools include:

• Styles and presets: Change the visual direction without rebuilding a pipeline.
• Image prompts: Guide a result with a source image.
• Inpaint and outpaint: Replace or extend selected parts of an image.
• Upscaling and variations: Refine an output without leaving Fooocus.
• LoRAs: Add a compatible SDXL style or concept.

Fooocus can work offline after installation and required downloads. Initial setup, models, inpaint components, and updates still need network access.

If Fooocus Returns a Black Image

A black output usually means the optional blackout path is enabled.

1. Turn off Black Out NSFW in the interface.
2. Check that default_black_out_nsfw wasn't set to true in your config.
3. Restart Fooocus and test again.
4. If you edited Python files using an old guide, restore the official files or reinstall cleanly.

Don't assume every black image comes from that setting. A broken model file, memory failure, or generation error can also prevent a valid output.

Common Problems

The model won't load

MetadataIncompleteBuffer and PytorchStreamReader errors usually point to a corrupted checkpoint. Delete that file and download it again from its original source.

Fooocus crashes with a memory error

The official troubleshooting guide connects RuntimeError: CPUAllocator and related failures to memory, swap, and free disk space. Keep virtual swap enabled and make sure the relevant drives have enough room.

A Flux checkpoint doesn't appear or run

Fooocus is SDXL-focused and has no current plan to add newer architectures. Use RuinedFooocus, Forge, ComfyUI, or SwarmUI for Flux.

An old guide says to edit the safety checker

Restore the official installation. The supported current behavior is controlled by the optional blackout setting, which already defaults to off.

The Limits of Fooocus

Fooocus is still one of the easiest local SDXL interfaces, but it is in limited long-term support. Expect bug fixes, not major new model families.

Use Fooocus when you want:

• A simple prompt-first SDXL workflow.
• An official Windows archive with automatic model downloads.
• Built-in inpaint, image prompt, upscale, variation, style, and preset tools.
• Local generation without a hosted prompt policy.

Pick another tool when you want:

• Newer architectures in a familiar interface: Read the RuinedFooocus overview.
• Reusable graph workflows: Use ComfyUI.
• A larger form-based WebUI: Use Forge.
• A paid preconfigured install: LocalForge AI is one option if you don't want to manage Python and models yourself.