Background Remover
Remove an image background with AI that runs entirely in your browser — nothing is uploaded.
Background Remover cuts the subject out of a photo using BiRefNet, an open-source AI segmentation model (MIT license), running entirely in your browser through ONNX Runtime Web — via WebGPU where available, WebAssembly everywhere else. Unlike typical background removers, there is no server in the loop: no upload queue, no API, no copy of your photo on anyone’s computer.
JPG / PNG / WebP — processed 100% on your device. Your image is never uploaded.
- The AI runs on your device. This page uses an open-source AI model (BiRefNet, MIT license) executed by ONNX Runtime Web (MIT) through WebGPU or WebAssembly. There is no background-removal server and no API key — the technology simply does not send your image anywhere.
- First use downloads ≈ 103 MB once — the AI model (≈ 91 MB, most weights compressed to 8-bit) plus the runtime (≈ 12 MB). Both are served from this website and cached by your browser, so repeat visits load in about a second and work even on a slow connection.
- Expected times: the one-time download takes roughly 10–60 seconds depending on your connection. Each removal then takes about 2–10 seconds on a desktop, and can take noticeably longer on older phones. The status line above always shows what is happening — a download never runs silently.
- Your image never leaves your device. Nothing is uploaded, stored, or logged. Close the tab and everything is gone. You can verify this in your browser's network inspector: after the model files, the only requests are this page's own assets.
- Honest limits: an on-device model is not a cloud studio. Fine hair, fur, glass and transparent objects may keep small fringes. Images are analysed at 1024×1024 internally, then the cut-out is applied to your image at its full original resolution.
How to use
- Drop an image (or click / paste one). On first use your browser downloads the open-source AI model (≈91 MB) from this site — watch the progress bar; it is cached afterwards.
- The model runs on your device and cuts out the subject. Pick a transparent, white, black or custom-color background, an output size, and whether the subject fills the frame or fits inside it.
- Download the result as PNG (keeps transparency) or JPG. Your image was never uploaded — close the tab and it is gone.
About this tool
The trade for that privacy is a one-time download: the model (served from this site with its weights compressed to 8-bit, ≈91 MB) plus the runtime (≈12 MB). Your browser caches both, so later visits start in about a second. Each removal takes a few seconds on a desktop and longer on older phones — the status line on the page always tells you what stage it is at.
After the cut-out you can keep the background transparent, or place the subject on white, black or any custom color, and export at the original size or common marketplace ratios (1:1, 4:5, 5:4, 16:9) — the everyday needs of a product photo, covered without an account or a watermark.
How it works
The model is a segmentation network: it reads the image (resized internally to 1024×1024 and normalized) and outputs a per-pixel score — a logit — for 'subject' versus 'background'. A sigmoid turns each score into an opacity between 0 and 1: that grid of values is the alpha mask. The mask is then scaled back up to your image’s original dimensions and multiplied with the original pixels, so the exported cut-out keeps full resolution even though the AI 'saw' a smaller copy.
Most weights served here are the open-source BiRefNet (lite) model, compressed from 32-bit floats to 8-bit integers per channel; a few precision-critical tables inside the decoder are deliberately kept at full precision, because compressing them measurably hurt quality. That brings the download from 214 MB to about 91 MB; before shipping this build we compared its masks against the full-precision model and measured agreement (IoU) of 0.97–0.99 on our test images, 0.98 on average. At runtime the values are restored (dequantized) and all math runs in 32-bit, on your GPU (WebGPU) or CPU (WebAssembly).
Solid backgrounds and resizing are plain canvas work after the cut-out: the subject is scaled with a contain fit (whole subject visible, background fills the margins) or a cover fit (frame filled, edges cropped), composited over your chosen color, and encoded as PNG (with alpha) or JPG (flattened).
Worked examples
Marketplace main image: A shoe photographed on a cluttered desk becomes a clean white-background 1:1 image — the format most marketplaces demand — without the photo ever leaving the seller’s laptop.
Profile picture with a brand color: Cut a headshot out, set a custom background to your brand color, export 1:1 — a consistent avatar set for a whole team page, no uploads to a third-party editor.
Quick reference
| Stage | What happens | Where |
|---|---|---|
| Model download (≈91 MB, once) | Browser fetches the AI model + runtime from verystep.com and caches them | Your device |
| Analysis | Image resized to 1024×1024, normalized, run through BiRefNet | Your GPU / CPU |
| Mask | Logits → sigmoid → alpha mask, scaled to full resolution | Your device |
| Export | Subject composited on chosen background/size, saved as PNG or JPG | Your device |
Frequently asked questions
Is my photo really never uploaded?
Really. The AI model runs on your device; the page makes no request carrying your image — you can confirm it in your browser’s network inspector, or disconnect from the internet after the model has downloaded and keep removing backgrounds offline.
Why does the first use take longer?
Your browser has to download the AI model (≈91 MB) and runtime (≈12 MB) once. On a typical broadband connection that is 10–60 seconds. They are cached on your device afterwards, so the next visit skips straight to processing.
How good is the quality compared with paid tools?
For clear subjects — products, people, pets — it is very good. Cloud studios still win on the hardest edges: flyaway hair, fur against busy backgrounds, glass and transparent objects can keep a slight fringe, because this model analyses the image at 1024×1024 (the cut-out itself is applied at your image’s full resolution).
Which model and runtime does it use?
BiRefNet (lite variant) in ONNX format, with weights compressed to 8-bit by us — validated to match the original model’s masks almost exactly. Runtime: ONNX Runtime Web 1.22. Both are open source (MIT). The model file is served from verystep.com itself, not a third-party CDN.
Is there a file size or resolution limit?
There is no hard limit, but very large photos (40+ megapixels) can run slowly or exhaust a phone’s memory. If a removal fails, the page says so — try resizing the image first with our Image Resizer.
More free tools
Image Resizer
Resize images in your browser — no upload.
Image Format Converter
Convert between PNG, JPG and WebP locally.
Image Compressor
Shrink image file sizes without uploading.
Image Resizer
Resize and crop images to standard dimensions — social media, passport, ID photos.
OCR Image to Text
Extract text from photos and screenshots — right in your browser, nothing uploaded.
EXIF Viewer & Cleaner
See the hidden data inside your photos — then strip it before sharing.