Limitations of browser-native processing

Concrete memory and size limits

• Photo decode memory: a 12-megapixel photo occupies roughly 48 to 50 MB of memory once the browser decodes it for canvas processing, regardless of the source file's compressed size on disk; a batch of 100 such photos can temporarily require several gigabytes.
• PDF tools: merge, split, and compress do not enforce a software file-count limit, but documents in the range of 100 MB or more, or batches of multi-hundred-page files, can exceed a browser tab's available memory during the load and re-save cycle.
• Diff checker hard guard: diff-checker.js enforces a combined character limit of 200,000 and a line-product limit of 250,000 (left lines times right lines); past either threshold the tool blocks the comparison outright with a message that the comparison is too large for the browser preview, rather than attempting it and risking a frozen tab.
• Format rejection: the images-to-pdf and metadata-remover tools validate every added file and explicitly count files marked invalid, reporting a rejected total in the status line instead of silently dropping unsupported files.

Background remover's AI mode

The background-remover tool's default mode is a local flood-fill that adds no external library. Its optional 'AI Deep Cutout' mode is different: selecting it loads the @imgly/background-removal library from jsDelivr and downloads an approximately 40 MB model file on first use, a one-time transfer specific to that mode.

In the privacy build, this optional mode is not silently disabled by omission; the code path that would fetch the library is replaced with logic that rejects the operation and throws an explicit error stating that AI background removal is not available in the privacy build. The tool's default, non-AI mode is unaffected and continues to run against canvas image data in both builds.

What the public site loads besides tool code

Privacy limitations

• Compromised device: a compromised browser, operating system, or malicious extension can observe page data; this is an explicit non-goal in THREAT_MODEL.md, not something the app can defend against.
• Host-level injection: the privacy deployment depends on the hosting platform not injecting runtime scripts; Cloudflare Web Analytics, Zaraz, and similar features must be disabled at the dashboard level for the privacy hostname.
• Downloaded files: once a file is downloaded or shared elsewhere, it leaves the page's control.
• Re-encoding fidelity: some conversions re-encode files and can change image or PDF fidelity; THREAT_MODEL.md does not claim every file format is fully sanitized after conversion.
• Visible content and metadata: canvas-based metadata removal strips common embedded metadata by re-encoding the image, but it does not remove faces, addresses, visible text, or other pixels, and is not validated against every high-risk publishing workflow.

FAQ

When should I avoid browser-native tools?

Avoid them when files are too large for your device, formats are unsupported, or you need automation/API workflows.

Does metadata removal remove visible private information?

No. It removes metadata by re-encoding supported images, but visible pixels remain.

Does the privacy build protect against extensions?

No. Browser extensions operate outside the app's control.

Does the privacy build include the AI background-removal mode?

No. That code path throws an explicit error instead of loading the model, so only the default flood-fill mode is available there.

What exactly triggers the diff-checker size guard?

Either the combined character count of both text boxes exceeding 200,000, or the product of left-side and right-side line counts exceeding 250,000.

Does the public site send my file names or file content to analytics?