Our story

It started with a black box in Photoshop

I was working on product photos — cutting out items to put on clean backgrounds — and I ran the images through one of the popular free tools. The preview looked great. The edges were sharp, the subject was cleanly isolated. I downloaded the PNG and dropped it into Photoshop.

Black background. Not transparency.

I assumed I’d done something wrong. But no — the tool had exported premultiplied alpha, which bakes the background color into the RGB values of every semi-transparent pixel. It looks fine in a browser. It looks broken everywhere else.

Every free tool had the same problem

I tested a dozen tools. The pattern was consistent: beautiful previews, broken exports. Some tools charged for the “high quality” tier, which turned out to mean higher resolution — not better alpha. Others sent your images to a server and kept them for “model improvement.”

The paid tools — remove.bg and similar — got it right. But paying per image for something a neural network can do locally, on-device, in seconds, felt wrong. The models are open-source. The compute is your GPU. The only thing you’re paying for is someone’s server costs and the pretense that this is somehow expensive.

So I built it

NSS Background Remover runs the entire pipeline in your browser. WebGPU for fast inference, WebAssembly as fallback, Transformers.js to load the models. The image never leaves your machine. There is no server.

The alpha is always straight. I wrote an automatic integrity check that decodes every export and samples edge pixels to confirm the transparency is real before the download starts. If it fails — which happens when an encoder silently premultiplies — the export is rejected and an error is shown.

The edge decontamination, the Float32 mask pipeline, the bicubic upscale with unsharp mask — all of it exists because I got annoyed at tools that did the easy version of a problem instead of the right version.

Why it’s free

The underlying models are open-source. The compute runs on your hardware. The only real cost is hosting the static site, which Vercel handles cheaply.

The tool is ad-supported — non-intrusive display ads help cover costs. That’s it. No freemium, no watermarks, no usage limits. Quality isn’t a premium feature here. It’s the whole point.

The Ricky test

We have an internal quality benchmark called the Ricky test: cut out a Yorkshire Terrier with the best quality model, export as PNG, open in Photoshop. If you see natural fur edges with no halo and no black outline around individual hairs — the tool is doing its job. If any change to the pipeline makes the Ricky test look worse, that change doesn’t ship.

It’s a specific, honest benchmark. Fur edges at 4K are the hardest thing to get right in background removal. If you can nail that, everything else is easy.