Remove an Image Background Without Uploading Your Photo

Background removal is the most-searched image edit on the web, and the dominant services all share one design: your photo goes to their servers, their AI processes it there, and the full-resolution result usually costs credits. That design made sense in 2019, when segmentation models needed datacenter GPUs. It no longer does.

The model now fits in your browser

The AI that separates a subject from its background is a segmentation model — it looks at every pixel and decides "subject" or "not subject." Open models like RMBG-1.4 do this remarkably well and weigh in around 44 MB. That is small enough to download once, cache, and execute in the browser through WebAssembly.

So the architecture flips, the same way it has for PDFs and transcription:

• Upload model: your photo (possibly 10+ MB, possibly private) travels to a server; a preview comes back free and the full resolution is metered.
• Local model: the 44 MB engine travels to you, once. Every photo afterwards is processed on your own hardware, at the original resolution, for free — because it costs the site nothing per image.

Our Background Remover runs this way, and shows a live counter of how many bytes were uploaded to the server. It reads 0 because it is 0.

What you give up, honestly

The big commercial services train proprietary models on enormous datasets, and on the hardest cases — flyaway hair against a busy background, translucent fabric, fine fur — they are still a step ahead. If you need a perfect cutout of a golden retriever mid-shake for a billboard, a paid service may earn its credits.

For the bread-and-butter cases — product shots, portraits, screenshots, logos, marketplace listings — the open model's output is hard to tell apart, and you can judge it side by side before downloading anything.

Note
The first run downloads the model (~44 MB) and takes a moment. Every run after that starts instantly from cache. That download is the engine, not your image — data flows one way, toward you.

A practical workflow

1. Run your photo through the Background Remover and download the transparent PNG.
2. If the source was small, pass the result through the Image Upscaler for a sharper 2× version — also local.
3. Photos from a phone carry GPS coordinates and device details; the EXIF Remover strips them before you publish.
4. Resize or crop for the destination with the Image Resizer.

Every step in that chain happens on your device. The only network traffic is the one-time engine download.

Why this matters more for photos than for most files

A document leaks what is written in it. A photo leaks more: faces, your home's interior, license plates, reflections, and — in the metadata — often the exact coordinates where it was taken. "Uploaded then deleted" is a weaker promise when the file itself is this identifying. Local processing simply removes the question.

We didn't leave that claim as marketing, either. In our Privacy Audit we ran this exact tool with the network inspector open and published what we captured: the only requests were the model files downloading to the browser — config.json and model.onnx, by name — and the photo we processed never appeared in any outbound traffic. You can reproduce that check yourself in about 30 seconds.

Getting a better cutout

The model does the work, but the input decides how much work there is to do. Three things that measurably improve results:

• Contrast between subject and background. A dark product on a light table segments almost perfectly; a gray cat on a gray couch makes any model guess. If you control the photo, control the backdrop.
• Resolution in the useful range. Extremely small images give the model too few pixels to find an edge; enormous ones just take longer. Anything from about 800px to 4000px on the long side is the sweet spot. If your source is tiny, upscale it after the cutout, not before — upscaling first also sharpens the background you're about to discard.
• One clear subject. The model keeps what it judges to be the foreground. Group shots and busy scenes work, but the cleanest results come from a single, well-lit subject.

Edges with fine hair or translucency are where you should zoom in and check before shipping the result — that's true of every background remover, paid ones included.

PNG or WebP for the result?

The cutout downloads as a PNG because PNG preserves the alpha channel (the transparency) losslessly — that's non-negotiable for a cutout. But a full-resolution transparent PNG can be heavy. If the destination is a website, run the result through the Image Converter to WebP, which also supports transparency at a fraction of the size. If it's headed for a slide deck, print, or a marketplace that composites on white anyway, the PNG is fine as-is.

One thing to avoid: converting a cutout to JPG. JPG has no transparency, so the alpha channel gets flattened onto a solid background — which quietly undoes the edit you just made.

Hitting marketplace and profile-photo specs

Most background removal ends up serving one of a few destinations, and each has its own demands:

• Marketplace listings (eBay, Etsy, Amazon-style) generally want the product on a plain or white background at 1000–2000px on the long side. Remove the background, then use the Image Resizer to hit the pixel spec exactly — most platforms compress whatever you upload, so giving them a clean, correctly-sized file is what keeps edges crisp.
• Profile photos and avatars are displayed small and often circular. Crop square with the Image Cropper with your face centered before resizing down; cropping after shrinking throws away sharpness you can't get back.
• Slide decks and documents are the easy case: the transparent PNG drops onto any slide background directly, no further processing needed.

The order matters more than people expect: cut out first, crop second, resize last, and convert format only at the very end. Every one of those steps stays on your device.