What Stays Local in LocalForge

Local generation in plain language

LocalForge runs the generation workload on your computer after the required engine and model are installed. In that workflow, the prompt, selected local model, generation settings, and output are processed by local hardware rather than sent to a hosted image-generation service.

That statement has a defined scope. It does not mean every file on the computer is automatically private, every linked website works offline, or the operating system has no network activity.

Prompts and settings

Prompts, negative prompts, seeds, dimensions, sampler choices, and other generation settings are inputs to the local generation job. Protect project files and saved presets with the same care as other creative documents, especially on shared computers.

Source images

Image-to-image, inpainting, and reference workflows use files you select. Keeping generation local avoids the cloud-generator upload step, but the original file may still exist in synced folders, backups, recent-file lists, or project exports. Check those storage locations when the source material is sensitive.

Models and supporting assets

Installed checkpoints, LoRAs, VAEs, embeddings, and related assets remain files on your storage. Their source, license, and integrity are separate concerns from privacy. A local file can still have restrictive licensing or come from an untrusted source.

Generated outputs

Outputs are written to local storage according to the application and project configuration. Local storage is not automatically secure storage. Review:

• The output folder and its permissions.
• Cloud-sync software watching that folder.
• Automated photo-library imports.
• Backup destinations and retention.
• Shared user accounts and removable drives.
• Metadata retained in exported images.

When a network may still be involved

First launch may download the engine; generation can run offline afterward. Downloading new models, viewing online documentation, contacting support, and obtaining updates are internet activities. They are not the same as uploading a prompt to a hosted generator, but they should still be included in a serious threat model.

A practical privacy verification

1. Complete setup and install one compatible model.
2. Generate a small connected baseline.
3. Close LocalForge and disconnect Wi-Fi and Ethernet.
4. Reopen the app and repeat the generation with the installed model.
5. Confirm the output is written locally.
6. Inspect the output folder for sync or sharing rules.
7. Reconnect only for a specific maintenance task.

Shared and managed computers

Local processing does not defeat administrator access, endpoint monitoring, malware, screen recording, or another user with filesystem permissions. Use an account and device that match the sensitivity of the work. Encrypt the device and backups when appropriate, and keep the operating system patched.

What this page does not claim

It does not claim complete network independence, anonymity, immunity from malware, or automatic legal rights to models and outputs. Privacy depends on the whole system: LocalForge, storage, user accounts, backups, network configuration, source assets, and operating-system controls. Runs on Windows, macOS, and Linux.

Document your boundary

For repeatable sensitive work, record which folders are local, which folders sync, who can access the account, which model files are approved, and when the machine is intentionally connected. A written boundary is easier to audit than a vague promise.