How to use Background Remover
- 1Drop in a photo — nothing is uploaded
- 2Pick Fast for people and portraits, or Best quality for products and animals
- 3Wait once while the model downloads, then it stays cached
- 4Fix any missed edges with the erase, restore and magic wand brushes
- 5Choose a transparent or coloured background and download the cutout
What Background Remover does
Removing a background means deciding, for every pixel, how much of it belongs to the subject. That is easy in the middle of a person and genuinely hard at the edges, where a single pixel may be part hair and part wall. Older tools approached this by colour — select everything green, delete it — which is why they needed a green screen to work at all.
Segmentation models learn the shape of a subject instead of its colour, so they work on ordinary photographs with cluttered backgrounds. What they produce is not a yes-or-no selection but an alpha mask: a value per pixel from fully transparent to fully opaque, which is what lets a cutout keep soft edges on hair and fur rather than a hard jagged outline.
When to use it
- Product photos for a listing
- Marketplaces overwhelmingly want a plain white background. Cutting the product out and setting white is faster and more consistent than reshooting against a backdrop.
- A profile picture that sits on any colour
- A transparent PNG drops onto a dark or light page without a visible box around it, which is exactly what a website header or a slide deck needs.
- Design assets
- A cut-out subject can be composited over anything. Because the mask keeps soft edges, hair and fur blend into the new background instead of showing a hard cut line.
- Removing a distracting setting
- A good photo taken in a bad place — a cluttered room, a busy street — becomes usable once the subject is separated from what is behind them.
Background Remover settings explained
- Fast
- MODNet, 24 MB, Apache-2.0 licensed. Trained on people, so it is the right choice for portraits and anything human. Downloads quickly and runs in a couple of seconds on most devices.
- Best quality
- BiRefNet Lite, 114 MB, MIT licensed. Handles products, animals and general subjects that MODNet was not trained for. Loaded at half precision, which keeps it at 114 MB instead of 213 MB with no visible difference in the mask.
- Erase and restore brushes
- Paint away parts the model kept, or paint back parts it removed. Adjustable brush size. Faster than re-running the model when only a small area is wrong.
- Magic wand
- Click a pixel to select everything similar in colour, with a tolerance slider. Contiguous mode limits it to the connected region under the cursor; turning it off selects matching colour anywhere in the image.
- Feather
- Softens the mask edge by a chosen number of pixels. A value of 1 or 2 removes the hard staircase you get on a diagonal edge; higher values help when compositing over a busy background.
- Background
- Keep the cutout transparent, or fill behind it with a colour. Transparent output must be saved as PNG or WebP — JPEG cannot store an alpha channel and will flatten it.
Which model to use
| Model | Download | Trained for | Licence |
|---|---|---|---|
| Fast (MODNet) | 24 MB | People and portraits | Apache-2.0 |
| Best quality (BiRefNet Lite) | 114 MB | Products, animals, anything | MIT |
Limits and what to expect
- The model downloads once per browser — 24 MB or 114 MB — then stays cached for later visits.
- Running the model needs enough memory for the image plus the network; very large photos may fail on older phones.
- Fine detail like flyaway hair, chain-link and glass is where every segmentation model struggles; the brushes exist for exactly that.
- Semi-transparent subjects such as glass or smoke cannot be separated properly — the model produces an outline, not true transparency.
- One subject at a time: the mask covers whatever the model considers foreground, and cannot be told to keep only one of two people.
Troubleshooting
- The edges look rough or stepped
- Raise the feather value to 1 or 2 pixels. That softens the transition and removes the staircase effect on diagonal edges. Higher values help when the cutout is going over a busy background rather than a flat colour.
- Parts of the subject were removed
- Switch to the Best quality model if you used Fast on a non-human subject — MODNet is trained on people and will misread a product or an animal. Then use the restore brush to paint back anything still missing.
- The transparent background turned white or black
- The image was saved as JPEG, which has no alpha channel, so transparency is flattened onto a solid colour. Export as PNG or WebP instead, both of which store transparency.
- The model is taking a long time on first use
- That is the one-time download — 24 MB for Fast, 114 MB for Best quality. It is cached afterwards, so later images are processed immediately. On a slow connection, start with Fast.
Your files and your privacy
The model is downloaded to your browser and the photo is processed on your own device — the image itself is never transmitted. This is a meaningful difference from hosted background removers, which upload your picture to their servers and typically retain it for some period. The only network request here is for the model weights, which are the same generic files for every user and carry nothing of yours.
What to use next
FAQ
Is the background remover really free?
Yes, with no account, no watermark and no daily cap. That is possible because the processing happens on your device rather than on a server — there is no per-image cost to recover, which is what drives the credit systems on hosted services.
Does my photo get uploaded?
No. The segmentation model is downloaded to your browser the first time you use it, and every image after that is processed locally. You can disconnect from the network once the model has loaded and the tool still works.
Which model should I choose?
Fast (MODNet, 24 MB) for people — it is trained specifically on portraits and is quick to download. Best quality (BiRefNet Lite, 114 MB) for products, animals and everything else. Using Fast on a non-human subject is the most common cause of a poor cutout.
Why is the first run slow?
The model has to download once, and then it is cached by the browser. After that, cutting out an image takes a couple of seconds. Choosing Fast rather than Best quality reduces that first wait from 114 MB to 24 MB.
Can I fix a bad cutout by hand?
Yes. The erase brush removes areas the model kept, the restore brush paints back areas it removed, and the magic wand selects by colour with an adjustable tolerance. Hair, fur and thin structures usually need a small manual pass whatever tool produced the mask.
What format should I save the result as?
PNG for a transparent background, or WebP for the same transparency at a smaller size. Do not use JPEG — it has no alpha channel, so the transparent area is filled with a solid colour when the file is written.