How VanishBG Keeps Your Photos Private (On-Device AI Explainer)

Most background-removal tools work the same way: you upload your photo to a server, the server runs an AI model on it, and the server sends back the cutout. The photo — and the right to do whatever the server’s owner wants with it — leaves your device.

VanishBG doesn’t work that way. The AI runs entirely inside your browser. Your photo never leaves your device. This post explains how that’s possible, what it actually means for your privacy, and why it matters.

What “on-device AI” means

“On-device AI” means the AI model — the mathematical function that takes your image as input and produces a cutout as output — runs on your phone or laptop’s processor, not on a remote server.

Concretely, when you drop an image into VanishBG:

1. Your browser downloads the AI model once (a one-time ~30–60 second download from a public CDN).
2. The model is cached by your browser’s HTTP cache.
3. When you process an image, the model runs locally using your device’s CPU and GPU via WebAssembly and WebGL.
4. The cutout is produced locally and displayed locally.
5. No image bytes are transmitted to any server at any point.

The only network calls are:

• The initial page load (HTML, CSS, JavaScript) — same as any website.
• The one-time model download — a static file from unpkg, a public CDN.
• (Optional) Cloudflare Web Analytics — cookieless, aggregate page-view counts only.

That’s it. There’s no API call with your image. No server-side processing. No storage. No logs containing your photo.

How this is technically possible

Until recently, running an AI segmentation model in a browser was impractical. Three things changed:

1. WebAssembly (WASM)

WASM is a binary instruction format that lets code written in languages like C++, Rust, and Python run in the browser at near-native speed. AI frameworks like ONNX Runtime can compile their models to WASM, allowing the same model that would run on a server to run in the browser with minimal performance penalty.

2. WebGL / WebGPU

AI inference is essentially a lot of matrix multiplication, which GPUs are very good at. WebGL has been available in browsers for over a decade, and modern implementations can run the kind of parallel computations AI models need. WebGPU, the next-generation API, is even better but isn’t yet universal.

VanishBG’s engine (built on @imgly/background-removal) uses WebGL when available, falling back to WASM-only on devices without GPU acceleration.

3. Better models, smaller sizes

The AI segmentation models used for background removal have gotten dramatically better and smaller over the past few years. The ISNet model that powers the default engine is around 25 MB — small enough to download once and cache, large enough to produce high-quality cutouts on a wide range of subjects.

Compare this to a few years ago, when the only viable models were hundreds of megabytes and required server-side GPUs to run in reasonable time.

Why this matters for privacy

When you upload a photo to a cloud-based tool, you’re trusting that company with:

1. The photo itself. Faces, documents, products, IDs — anything in the image.
2. Metadata about you. Your IP address, browser fingerprint, time of upload, account (if you have one).
3. The right to use the photo. Read the terms of service of most cloud tools — many claim broad licenses to use your uploaded content for their own purposes, including training future AI models.

Even if the company is trustworthy today, terms of service change, companies get acquired, data breaches happen, and “anonymized” datasets get re-identified. Once your photo is on someone else’s server, you’ve lost control of it.

On-device AI eliminates all of these concerns by design. There is no server with your photo. There is no log of what you processed. There is no license you’ve granted. The transaction between you and the AI happens entirely on hardware you control.

Specific scenarios where this matters

E-commerce sellers with unreleased inventory

If you’re photographing a product that hasn’t been announced yet, uploading it to a cloud-based background remover means trusting that company not to leak or use the photo. For unreleased tech products, fashion collections, or limited-edition releases, a leak can cost real money.

With on-device AI, the photo stays on your laptop. No risk of pre-release leak.

Journalists and activists

Photos of sensitive subjects — sources, documents, locations — used to require trusting a cloud service. If you’re working on a story the government or a corporation wouldn’t want published, that trust is a vulnerability.

On-device AI removes that vulnerability. You can process sensitive photos without revealing them to anyone.

Personal photos

Family photos, medical photos (skin conditions, injuries), intimate photos — all of these are routinely uploaded to background-removal services without much thought. The privacy implications are significant.

On-device AI keeps these photos private by default.

Photos of children

Many parents are careful about what they post on social media but think nothing of uploading their kids’ photos to a “free tool.” On-device AI is the safer choice.

Photos containing confidential information

A screenshot of a slide deck, a photo of a whiteboard, a picture of a document — these often contain information you wouldn’t want to share. On-device AI processes them without exposing them.

What on-device AI doesn’t protect against

On-device AI is not a complete privacy solution. It protects your image bytes from being transmitted, but it doesn’t protect against:

• Your browser being compromised. Malware, malicious extensions, or compromised browsers can read what’s in memory.
• Your device being seized. If someone has physical access to your device, they can recover cached data.
• Network-level surveillance of the page load itself. Your ISP can see that you visited vanishbg.com, even if they can’t see what you did there.
• Cloudflare logging. Cloudflare (which serves the site) logs standard request metadata for security purposes. This doesn’t include your image, but it does include your IP address and the time of your visit.

For most users, these limitations are theoretical. For users facing serious threats (journalists in authoritarian countries, dissidents, etc.), additional measures like VPNs, Tor, and hardened browsers are appropriate.

How to verify the claim

The “on-device” claim is testable. Here’s how to verify it yourself:

1. Open VanishBG in your browser.
2. Open DevTools (F12) → Network tab.
3. Process an image.
4. Watch the network requests. You should see:
   • The initial page load (HTML/CSS/JS)
   • One large download the first time you process (the AI model, from unpkg.com)
   • Cloudflare Web Analytics beacon (cookieless, aggregate)
5. You should NOT see any request containing image data going to a VanishBG-controlled server.

If you see image data being uploaded, that’s a bug and we want to know about it.

The future of on-device AI

The shift from cloud-based AI to on-device AI is one of the most important privacy trends of the 2020s. As models get smaller and browsers get more capable, more and more AI-powered tools will run entirely on the user’s device.

This is good for users (privacy, no per-use cost, works offline) and good for tool builders (no server costs, no API rate limits, no data compliance headaches). The main losers are the cloud AI companies that built business models on aggregating user data — but that’s a feature, not a bug.

VanishBG is one small example of this shift. As the ecosystem matures, expect to see on-device AI for translation, transcription, image editing, code completion, and many other tasks that currently require server roundtrips.

Bottom line

When you use VanishBG, your photo never leaves your device. The AI runs locally, the cutout is produced locally, and the only network traffic is the one-time model download. This isn’t a marketing claim — it’s a verifiable technical fact, and it’s the single biggest reason to choose an on-device tool over a cloud-based one.

If you’ve been uploading sensitive photos to cloud-based background removers, give VanishBG a try. Same quality cutouts, none of the privacy concerns.