Removing a background without the halo
The model finds the subject in a second. The last five percent — the edge — is where cut-outs are won and lost, and it has a specific, fixable cause.
7 min read · Updated
Private — processed on your device, never uploaded
Where the fringe comes from
Put a cut-out on a white background and you often see a dark rim around the hair and shoulders. It looks like the model got the outline wrong. It did not — the outline is fine, and the problem is one step further on.
A pixel at the edge of a subject is not purely subject or purely background. It is a blend of both, because the sensor averaged whatever fell across it. A strand of hair might cover 40% of a pixel; the other 60% is the wall behind. The model correctly marks that pixel as 40% opaque — but the colour stored there is still the mixture, wall included.
Composite that onto a new background and you are laying down 40% of a colour that is part old wall. Shot against something dark, every edge pixel carries a little darkness with it, and the sum of all of them is the halo.
The photo decides most of the outcome
No amount of edge refinement rescues an image where the subject and the background are the same colour and tone. If you have any control over the source, this is where to spend it.
- Contrast between subject and background matters more than resolution. A dark coat against a dark sofa is a hard cut-out at any megapixel count.
- Even lighting on the background prevents the estimator having to guess at several different background colours at once.
- Distance from the background reduces colour spill onto the subject in the first place.
- Avoid motion blur at the edges. A blurred edge is genuinely ambiguous, and nothing can un-blur it.
What the edge controls actually do
| Control | What it does | Reach for it when |
|---|---|---|
| Shrink | Pulls the edge inward by a pixel or two, discarding the most mixed pixels entirely. | A thin rim persists after decontamination. Costs fine detail, so use sparingly. |
| Feather | Softens the transition from opaque to transparent. | The cut looks unnaturally sharp against the new background, like a sticker. |
| Despill | Solves the old background colour back out of partly transparent pixels. | A coloured fringe — this is the control that removes a halo rather than hiding it. |
The order to try them: despill first, because it addresses the actual cause. Only then a pixel of shrink if a rim survives, and feather last and lightly. People usually do this backwards, feathering a coloured halo until it becomes a soft coloured halo.
Background removerAll three, as live sliders — the refinement runs on the finished cut-out in milliseconds, so you can see the effect while dragging.Check it against the background you will use
A cut-out inspected on the checkerboard almost always looks perfect. The checkerboard is mid-grey, and mid-grey hides both dark and light fringing. Judge the edge on the actual background — and if you do not know what that will be, judge it on both black and white, which is the worst case in each direction.
- View at 100%. Halos are invisible in a thumbnail and glaring in a full-size composite.
- Look at the hair, then the shoulders, then anything with a hard straight edge. Those three tell you almost everything.
- Save as PNG or WebP. JPEG has no transparency, and exporting a cut-out to JPEG silently flattens it onto white.
Questions
Why does my cut-out have a dark or coloured outline?
Because partly transparent edge pixels still hold the colour of the old background mixed in. The transparency is right; the colour is contaminated. A despill or decontamination setting solves the original foreground colour back out — shrinking or feathering only hides it.
Why does the hair look chewed?
Usually shrink set too high. It removes the most mixed pixels, and on hair those mixed pixels are the hair. Reduce shrink and lean on decontamination instead; some detail loss around fine strands is unavoidable if the contrast was low.
Which format should I save a transparent image as?
PNG for anything with sharp edges or text, and for maximum compatibility. WebP for photographic subjects where file size matters — it supports transparency and is usually smaller. Never JPEG: it has no alpha channel and will flatten the cut-out onto a solid background.
Does the background remover upload my photo?
It does not have to. The segmentation model can be downloaded once and run in your browser, which means the photo itself is never transmitted. The trade-off is the initial model download — a large download followed by no further network activity is what local processing looks like.
Private by design — your files never leave your device. If this was useful, passing it on is the whole marketing budget.