Browser AI, privacy, and offline checks
Follow how images, models, and cached files move through SansBG. Check network requests, local history, GPU fallback, and offline availability.
Updated · Maintained by Arunvpp
SansBG’s editor decodes your chosen image, calculates the cutout, and encodes the download in your browser. That local image-processing path still needs web resources to load, and the site can make requests unrelated to your image. This guide separates those activities and gives you checks to repeat on your own browser.
What happens to a selected image
- The browser opens the local file and draws a working copy to a canvas. Large images are reduced to the device profile’s size limit.
- Fast or Best quality loads an ONNX model and runs it in a worker. The model estimates a subject mask; another worker applies it to the working image. Plain background instead applies color or brightness rules and does not load an AI model.
- Touch up applies your Erase and Restore strokes locally. Export composites the selected background, size, crop, and shadow, then encodes a downloadable file.
The processing implementation does not submit selected image bytes to an inference server. Browser AI uses ONNX Runtime Web: WebGPU can run supported work on the graphics processor, while WebAssembly provides the CPU path. SansBG attempts a CPU fallback if GPU initialization or inference fails. Availability depends on browser, hardware, and runtime support.
Share and Copy are separate actions you initiate after processing. A receiving app or service has its own handling of a file you choose to share. Local processing does not make the downloaded file safe from other people using your device or account.
What can still use the network
| Resource | Why it is requested | What the host can receive |
|---|---|---|
| Pages, scripts, fonts, and sample images | Loads the site and any sample you choose. | Ordinary request information, such as IP address, resource URL, and browser headers. |
| AI model files | First use of a selected mode when its model is not cached. | A request for the model file, including normal connection and request information. |
| ONNX runtime files | Loads the code needed for the selected execution backend. | Ordinary resource-request information; downloading an engine is distinct from sending it an image. |
| Advertising and consent resources, when enabled | Supports the site’s advertising configuration and privacy choices. | Request and device information according to the provider and consent state described in the Privacy Policy. |
Model sources depend on the site configuration and available fallbacks. The code supports the site’s own model storage and pinned Hugging Face sources. Production WebGPU runtime downloads can use the site’s storage or jsDelivr. Downloaded models and the production WebGPU runtime are checked against expected hashes before use on secure origins. The fallback chain means that one blocked source may lead to a request to another host.
Read the Privacy Policy for the site’s current providers and advertising disclosures, and use Privacy and cookie settings in the editor page footer where available. Local inference does not mean there are no network requests or that third-party scripts receive no browsing information.
Repeat a network check with a harmless image
- Open a desktop browser’s developer tools and select Network. Keep recording enabled. Choose a non-sensitive test image with an easily recognized filename.
- Reload the page, then inspect the initial page and resource requests. Note whether advertising or consent resources are active in this session.
- Clear the displayed request log, select the local test file, and run Plain background. Inspect requests for their URL, method, initiator, and any request payload. The local file may appear as a
blob:resource; that is not an HTTP upload. - Repeat with Fast, then Best quality if your device can run it. An uncached mode can download an
.onnxmodel and runtime resources. A cached run can show much less network activity. - Inspect any unexpected request body for file bytes or a filename. Do not rely only on looking for POST: examine the complete request details and account for background page activity.
Chrome’s Network panel reference explains recording, request payloads, and preserving logs. A check describes that browser session and deployed version; it does not certify every future version or third-party script. If you report a surprising request, share its host, time, mode, and resource type without including private request contents.
What stays in this browser
Model files use the browser Cache API. The service worker separately caches visited pages and application resources for reuse. A preference remembers the selected AI mode, and GPU failures can be remembered for the current browser version. These are separate from Recent downloads.
When enabled, Recent downloads stores up to 12 single-image results locally in IndexedDB. Entries older than 30 days are removed when history is next viewed or a result is saved. Turning it off also clears saved history; Clear removes the current entries. Batch results stay in the open panel and do not enter that history.
Another person using the same browser profile may be able to view the saved results. Download files you need and clear history on a shared device. Browser storage is not a backup: storage quotas and eviction, private browsing, or clearing site data can remove cached models and results.
Prepare and test offline availability
- While online, process a harmless image in the mode you intend to use. Wait for the model and runtime to finish loading. Visiting the homepage alone does not load both AI models.
- Reload while still online so the service worker can control the page and cache the application files used by that session. Save a sample locally if you want to use it during the test.
- Disconnect the device, reload the same page, and choose a local file. Run the intended mode and download the result. Test a reload as well as an already-open tab.
- Repeat for another mode only after preparing it online. If an offline run fails, reconnect and identify the missing resource rather than assuming the image itself requires an upload.
Offline availability depends on the browser retaining the required page, scripts, runtime, and selected model. A successful test on one device does not establish support on another. Some cached assets may still try to refresh, and online browsing can contact hosting and third-party services. Save important outputs as files before a trip or a storage cleanup.