What local AI image generation really means.

In cloud image generation, a prompt or source image travels to someone else’s computers, the model runs there, and the result returns over the network. Local generation changes the center of that workflow: the model runs on hardware you control.

Local inference is the moment the model works

Inference is the computation that turns a prompt into an image or transforms an existing image. In Loccra, that computation runs through a local engine bound to your Mac’s loopback interface and uses Apple Silicon for acceleration. The creative request is not submitted to a public inference API.

That boundary matters because prompts can reveal product ideas, campaign plans, client names, private locations, or simply work that is not ready to share. Source images may be even more sensitive. Local inference keeps those inputs out of a remote generation queue.

The initial setup still needs a network

A useful image model is too large to hide inside a normal small app download. Loccra therefore needs an explicit one-time model installation. During this setup, it contacts the approved model source, downloads pinned files, verifies their checksums, and installs them into Loccra’s managed model directory.

This is different from sending prompts during generation. Network access during setup obtains the software resources needed to run locally; after setup, the creative operation can run offline.

• Setup: internet is required to download model data.
• Create: the prompt is processed locally.
• Edit: the source image and instruction are processed locally.
• Gallery: projects and result references are read from local storage.

What stays on the device during creation

During ordinary generation and editing, Loccra keeps the core creative data on your Mac:

• The positive prompt and optional Avoid text.
• An imported source image used for an edit.
• Generated image pixels.
• Seed, style, format, model identifier, and workflow metadata.
• The persistent project record used by Gallery.

Loccra also avoids logging full prompts by default. Diagnostics should describe technical failures without turning private creative content into a support log.

Where the work lives

The app uses managed directories in your user Library for models, outputs, projects, logs, and runtime data. Gallery points back to those local results so a creation remains available after relaunching the app.

Local storage gives you control, but it also makes disk planning important. Model files require meaningful free space, generation outputs accumulate over time, and macOS may need additional room for temporary files and memory swap.

Privacy comes with honest tradeoffs

A remote service can pool large GPUs and hide setup from the user. Local generation uses the memory, storage, thermals, and performance available on one Mac. The first model load can take time. Larger images need more unified memory. A 16 GB Mac should start conservatively, while a 32 GB Mac has more room for larger output.

The benefit is not that local computation is free; it is that the cost buys a different kind of control:

• No per-image inference upload.
• No account required for the creative loop.
• Offline work after setup.
• Predictable model selection instead of a changing cloud endpoint.
• Projects that remain part of your local workspace.

How to verify the promise in practice

After model setup, disconnect Wi-Fi and generate a small 640 px image. Then import a non-sensitive test image and perform an edit. Both should complete, appear in Gallery, and remain after relaunch.

You can also open Activity Monitor while generating. The local Loccra backend and its memory usage should be visible on your Mac. When Loccra quits, the backend process launched by the app should terminate rather than remaining as an exposed service.