🪄 Background Remover

AI background removal that never uploads your photo — the model runs in your browser. Transparent PNG out, no credits, no watermark.

About the Background Remover

Every mainstream background removal service works the same way: your photo is uploaded to their servers, processed there, and the result is sent back. That is a privacy trade most people never consciously made. This tool inverts it — the AI model comes to you, runs in your browser on WebGPU, and your photo never crosses the network at all. remove.bg cannot say that sentence; this page can.

The model is BRIA AI's RMBG-1.4, a purpose-built segmentation network. It downloads once (~44 MB, from the Hugging Face CDN), is cached by your browser, and produces a PNG with a real alpha channel. There are no credits and no watermark, because there is no per-image cost to recover: your hardware does the work.

How to use it

Drop a photo — people, products and animals segment best.
Click Remove background. The first run downloads the model; the page shows the size and progress rather than a silent spinner.
Inspect the edges with the before/after divider, especially hair and translucent areas.
Download the transparent PNG — compress it to WebP with the Image Compressor if it's headed for the web.

Frequently Asked Questions

Is my photo really not uploaded anywhere?

Really. The segmentation model (RMBG-1.4) is downloaded TO your browser and runs there — on your GPU via WebGPU where available, otherwise in WebAssembly. The only network traffic is the model itself, coming down once. Cloud services like remove.bg necessarily receive your photo to process it; this tool structurally cannot.

Why is there a ~44 MB download the first time?

That's the neural network doing the work. It comes from the Hugging Face CDN on your first removal and is cached by your browser, so later visits skip the download entirely. On a metered connection, wait for Wi-Fi — the button tells you the size before anything starts.

How good are the results?

RMBG-1.4 is a state-of-the-art open segmentation model: excellent on people, products and animals against reasonably distinct backgrounds — the e-commerce and profile-photo cases. Fine hair, glass and motion blur remain hard for every tool in this class, including the paid ones. The before/after slider shows you exactly what you got.

Why does my browser say it will be slow?

Without WebGPU the model runs on the CPU via WebAssembly, which takes roughly 15–30 seconds per image instead of a few. Chrome and Edge have WebGPU on by default; Firefox and Safari are still rolling it out. The tool warns before starting rather than leaving you staring at a spinner.

What format is the result?

PNG with a genuine alpha channel — the checkered area is real transparency, not white. It composites cleanly over any background. If you need a smaller file for the web afterwards, run the result through the Image Compressor and choose WebP, which keeps transparency.

Can I use the result commercially?

Your photo is yours, and processing happens on your machine. Note that the RMBG-1.4 model itself is released for non-commercial use by BRIA AI; for commercial pipelines, check their licence terms. This tool is free either way — there are no credits because there is no per-use cost.

Background removal in the browser: how it works and why local matters

Every mainstream background removal service — remove.bg, Canva, Photoroom, the feature inside every design app — works by uploading your photo to their servers, running a segmentation model there, and sending the cut-out back. The results are good; the arrangement is what deserves a second look. Your photos, often of people, travel to a third party under terms almost nobody reads.

This tool inverts the arrangement: the model comes to you. BRIA AI's RMBG-1.4 — a purpose-built background segmentation network — is downloaded to your browser (~44 MB, once, cached thereafter) and runs on your own hardware via WebGPU, with a WebAssembly fallback where GPU acceleration is unavailable. The photo never crosses the network. That sentence is the entire product, and it is one the upload services cannot say.

Because there is no per-image server cost, there are no credits, no watermark, no resolution ransom where the free tier returns a thumbnail. The costs are honest and stated up front: a one-time model download with a progress bar, and 15–30 seconds of CPU time if your browser lacks WebGPU (a few seconds with it).

The output is a PNG with a real alpha channel, previewed against the original with a draggable divider — because segmentation quality is best judged by your eyes on your image's edges, especially hair, fur and glass, which remain the honest weak spots of every model in this class, paid ones included.

How it works

On first use, transformers.js downloads the RMBG-1.4 ONNX weights from the Hugging Face CDN and caches them in your browser's Cache API keyed by model revision — the second visit skips the download entirely. The page shows the size and per-file progress rather than a silent spinner.

Your image is scaled to the model's 1024×1024 input, normalised, and run through the network — on the GPU via WebGPU where the browser supports it, otherwise on the CPU in WebAssembly. The output is an alpha matte: a per-pixel estimate of foreground probability.

The matte is resized back to your image's dimensions and written into the alpha channel of the original pixels on a canvas, preserving full original resolution and colour — the model decides transparency, never repaints content. The composite exports as PNG, the format whose alpha channel every tool respects.

Common uses

Cut out product photos for e-commerce listings without sending your catalogue through a third-party service.
Make profile pictures and team headshots with clean transparent backgrounds for sites and slide decks.
Remove cluttered backgrounds from photos of items for resale listings — pair with the EXIF stripper for location privacy.
Prepare cut-outs for marketing collages, thumbnails and hero images without a designer round-trip.
Process photos of people — kids' school projects, family prints — where uploading to an unknown server is exactly the thing to avoid.
Batch through a set of photos one by one: after the first image, the model is warm and each cut-out takes seconds.
Generate stickers and emotes from photos for chat apps, which accept transparent PNGs directly.

Before you rely on the result

Inspect edges with the divider before shipping — hair, fur, glass and motion blur are where every segmentation model earns or loses its keep.
Use photos with reasonable subject/background distinction; the model is excellent on people, products and animals, weaker on camouflage-grade scenes.
On a metered connection, wait for Wi-Fi before the first run — the model download is ~44 MB and the page tells you before starting.
If your browser lacks WebGPU (the page detects and says so), expect 15–30 seconds on CPU — it works, it is just not instant.
Compress the resulting PNG (WebP keeps transparency) before publishing — full-resolution PNG cut-outs are large.
Composite the cut-out over both a light and a dark background to catch faint halos before they ship.
For commercial pipelines, check BRIA's RMBG-1.4 licence terms — the model is released for non-commercial use; your photos, of course, remain entirely yours.