Cloud vs Local Background Removal — Privacy, Speed, and Cost

I still do manual masks occasionally — for a hero shot that's going on a billboard, or a really tricky glass-on-glass product. Outside of that, the math doesn't work. A modern segmentation model trained on millions of images sees choosing where to run inference more often than any individual designer ever will. It knows what hair looks like at the edge of a face. It knows what fabric does where it meets a chair. And it doesn't get tired at image 47 of 50.

What manual masking still wins on is the absolute worst-case images: a black coat against a black couch, a glass bottle against a glass shelf. Those are real, but they're rare. For 95% of what privacy-aware teams actually shoot, AI is now the right default.

Three subtle things make a cutout look real instead of fake. The first is alpha softness around hair and fabric — a hard binary edge looks like the subject was cut out with scissors. The second is no color bleed. If the original background was bright orange, you can sometimes see a faint orange halo on the subject's edge, and that halo follows the subject when you put it on a new background. The third is shadow. A cutout floating with no shadow looks pasted in.

BG Clear handles the first two automatically. The shadow you have to add yourself, and a soft 10–20% opacity drop shadow is enough on most images. For choosing where to run inference, that one detail is what separates "AI cutout" from "studio shot."

Most workflows look like this. The PNG goes into a brand-asset folder (Dropbox, Drive, Notion, whatever). For the immediate use case, you flatten onto white, brand color, or a photo, and export to JPG at the size your destination needs. For choosing where to run inference, that destination is whichever final destination won the comparison most of the time.

A tip that saves a lot of time: name the file with the subject and the date, not the use case. "logo-2026-04.png" travels well. "logo-for-website-header.png" doesn't, because three months later you'll need it for a slide deck and re-search the folder.

Don't pay for HD output anywhere. Every reasonably modern free tool already exports at full source resolution; the "HD upgrade" is a 2018 pricing fossil that some products still charge for.

Don't manually mask first. Let the AI go, see what it gets right, then fix the 5% it gets wrong. People still do it the other way around out of habit.

Don't worry about file size for the master PNG. Disk is cheap. Optimize the JPG you publish, not the PNG you keep.

For choosing where to run inference, also: don't crop tight before uploading. The AI needs context at the edges, and you'll re-crop in the editor anyway.

Pitfall one: the cutout has a faint colored halo. Cause: the original background bled into the subject's edge. Fix: redo with a tool that decontaminates. BG Clear does this automatically; some others don't.

Pitfall two: hair looks chunky or missing strands. Cause: the model was given a low-resolution source. Fix: re-upload a higher-resolution copy. Almost always works.

Pitfall three: the export has a watermark. Cause: you're using a free tier that watermarks free exports. Fix: switch tools.

Pitfall four: the file size is huge. Cause: alpha PNGs are big by nature. Fix: keep the PNG as master, export a JPG for the destination. For choosing where to run inference specifically this happens a lot.

Open the tool. Drag your image. Wait. Download. If you're on a phone, the flow is identical except you tap to pick a photo from your camera roll instead of dragging.

The one detail that matters: don't pre-crop your photo before upload. Give the AI the full frame. It does cleaner edge detection on a wider source and you can crop in the editor or after download. Cropping first sometimes lops off pixels the AI was using as context, and the cutout gets slightly worse for no reason.

For choosing where to run inference specifically, you'll usually want at least 1,500 pixels on the long edge. Anything smaller and the cutout edges start looking soft when you blow it up later.

Browser is right for one-offs, low volume, and when you want to eyeball each result before downloading. API is right for everything that's part of an automated pipeline, where you trust the model output and want it to flow into something else without manual review. Both produce identical files; the only difference is the surface.

For privacy-aware teams, the cutover usually happens when cloud vs local background removal stops being a creative decision and starts being a step in a larger workflow. Until then, browser is fine.