LocalForge AILocalForge AI

Fooocus Offline: Stop Update/Preset Fetches After Prep

Configuring Fooocus for Offline Use

Fooocus offline is possible: Fooocus AI needs the internet for first-time setup, then can run without it once the files are on disk. So do the downloads first, while you are still online, and disconnect only after that. Start each preset you plan to use once before going offline, because a preset fetches its own models the first time it runs.

What Fooocus downloads, and what to fetch before going offline

Everything below has to be on disk before the network goes off.

  • The package. On Windows, unzip the Fooocus package from the project's GitHub releases, then start run.bat.
  • The base model. The first run of run.bat downloads the base model, JuggernautXL V8 + Rundiffusion, which is 6.6 GB. Let it finish.
  • Preset models. The anime and realistic presets download their own models, but only the first time each one is started. Launch each preset you want while you are online.
  • A checkpoint you already have. Copy it into Fooocus/models/checkpoints and Fooocus has no need to download one.
  • The first generation. It is slow because of a lazy load and a prompt-expansion model download. Generate one image while online so that download is already done.

On Linux and Mac the steps differ from the Windows ones above, so follow the project's own instructions for your system.

Check that Fooocus really runs offline

Run this short test once the downloads are done:

  1. Disconnect the machine from the network (turn off Wi-Fi or unplug the cable).
  2. Launch Fooocus the way you normally do.
  3. Generate one image with a short prompt.

A pass looks like this: Fooocus opens, the image is generated, and the terminal shows no download or connection errors. If it stalls at launch or tries to download a model, reconnect, let that download finish, and run the test again.

Troubleshooting Offline Application Errors

A launch failure on an RTX 50-series card is not an offline problem. The error reads "no kernel image is available for execution", or "NVIDIA Geforce RTX 50** with CUDA capability sm_120 is not compatible with the current PyTorch installation". It happens because the PyTorch version Fooocus pins predates Blackwell (sm_120) cards. Reconnecting will not fix it, so check your GPU family first.

If a model download starts or fails when you launch with the network off, that file was not fetched beforehand. Go back to the list above, fetch it while online, and test again.