Background removal has stopped being hard, except on hair, glass and anything the same colour as what is behind it.
What the model is actually doing
It is not detecting edges or picking colours the way an older "magic wand" did. It is a segmentation network trained on a large number of images with the subject already marked, and it predicts, for every pixel, the probability that the pixel belongs to the foreground.
That prediction is why results are so uneven. The network is confident about shapes it has seen thousands of times — a person, a shoe, a bottle, a dog — and much less confident about anything unusual, or any region where foreground and background look alike.
It also explains why the failures look the way they do. You do not get a jagged edge; you get a soft, slightly wrong boundary, or a chunk of background retained because the model thought it was part of the subject.
What works and what does not
| Subject | Result | Why |
|---|---|---|
| Person, plain background | Excellent | The most heavily trained case |
| Product on white | Excellent | High contrast, hard edges |
| Pet or animal | Good | Fur edges soften slightly |
| Long or flyaway hair | Mediocre | Strands are thinner than the model resolves |
| Glass, bottles, jewellery | Poor | Transparency has no single right answer |
| Subject matching the background | Poor | Nothing distinguishes the boundary |
| Motion blur | Poor | The true edge does not exist |
| Several overlapping people | Variable | Often keeps only the most prominent |
If your subject is in the bottom half of that table, no amount of retrying will fix it. Either shoot again against a contrasting background, or accept a manual cutout.
How to get a better result
Most of the quality is decided before the tool runs.
- Contrast the subject against the background. A dark jacket on a dark sofa is the single most common failure. A plain wall of a clearly different tone solves most problems for free.
- Fill the frame with the subject. A person occupying 15% of a wide shot gives the model far fewer pixels to work with. Crop in first with the Image Cropper — a tight crop before removal beats a tight crop after.
- Use even light. Hard shadows read as part of the subject surprisingly often, and a shadow attached to a cutout looks worse than no cutout.
- Avoid busy backgrounds. Foliage and patterned wallpaper produce fragments the model keeps.
- Do not pre-compress. JPEG artifacts around the subject blur exactly the boundary the model is trying to find.
Saving it correctly
The most common way to waste a good cutout is to save it as JPEG. JPEG has no alpha channel, so the transparent area is filled — usually with black, sometimes white — and the transparency is gone permanently.
- PNG — universal support, full alpha, larger files. The safe default.
- WebP — transparency plus much better compression, typically 30% smaller. Good for the web, less good for handing to someone who will open it elsewhere.
- JPEG — only if you are placing the subject onto a new solid background and no longer need the cutout.
If the result will sit on a coloured background anyway, composite it first and then save as JPEG. That gives you a small file with none of the halo you get from a transparent PNG placed over an unexpected colour.
Why the first run downloads so much
A background remover that runs in your browser has to fetch the segmentation model before it can do anything — around 42 MB here. That is a real cost and worth explaining rather than hiding.
The alternative is uploading your photograph to a server that already holds the model. That is faster the first time and means the image leaves your device, which for portraits, product shots under embargo and anything with a person in it is a meaningful difference.
After the first run the model is cached and subsequent images are immediate. The Background Remover takes that trade deliberately: one slow start, then everything local and unlimited.
Frequently asked questions
Why does background removal fail on hair?
Individual strands are thinner than the model resolves, so it has to guess where the boundary sits and produces a soft, slightly wrong edge. Tied-back hair against a contrasting background works far better than flyaway hair against a busy one.
What format should I save a cutout in?
PNG for universal support, or WebP for the same transparency at roughly 30% smaller. Never JPEG — it has no alpha channel, so the transparent area gets filled with solid colour and the cutout is destroyed permanently.
Why does it keep part of the background?
Because the model judged those pixels to be part of the subject, which usually happens when they are similar in colour or tone to it. A dark subject on a dark background is the classic case. Shoot against a contrasting background if you can.
Can it handle glass or transparent objects?
Poorly. Transparency has no single correct answer — how much of the background should show through a bottle is a judgement, not a fact. These almost always need a manual cutout.
Why is the first use so slow?
A browser-based remover downloads the segmentation model once, about 42 MB here, then caches it. Server-based tools skip that by keeping the model on their hardware, at the cost of your photograph being uploaded to it.
Is there a limit on image size?
Not one imposed by a server, since the work happens on your device. Very large images use proportionally more memory, so a 50-megapixel photograph on an older phone may struggle where a laptop will not.