How to Remove an Image Background Cleanly

Automatic background removal works by running the image through a neural network trained to output a mask — a per-pixel judgement of how much each pixel belongs to the foreground. Everything downstream, whether you get transparency, a solid colour or a blur, is compositing applied to that mask. So the quality of a cutout is almost entirely the quality of the mask, and the things that make a mask hard are specific and predictable.

The subjects that work well share one property: the boundary between subject and background is unambiguous. A person against a plain wall, a product on a sweep, a logo on white. The model finds a clean edge because there is a clean edge to find.

The hard cases are the ones where the true boundary is genuinely not a line. Hair, fur, foliage, motion blur, smoke, glass and anything semi-transparent all occupy pixels that are honestly part subject and part background. A mask can represent that with partial values, but the finer the structure the more it becomes a guess — which is why the same tool that cuts out a bottle perfectly leaves a halo around curly hair.

The image is resized to the network's input size, normalised, and run through a segmentation model that outputs a single-channel mask at that resolution. The mask is then scaled back up to the original dimensions and used as an alpha channel. That sequence explains most of the artefacts people see.

Because the model works at a fixed, fairly small input size, fine structures thinner than a pixel at that scale cannot be represented — individual hairs, a chain-link fence, the teeth of a comb. Upscaling the mask afterwards smooths what it did capture, which is why fine detail comes back soft rather than sharp.

It also means image resolution beyond a point does not improve the cutout. A 24-megapixel photo and an 8-megapixel crop of the same subject produce very similar masks, because both are reduced to the same input size. What does improve the cutout is the subject being larger in the frame — cropping in before removing the background genuinely helps, where simply having more pixels does not.

Segmentation models are trained on datasets, and a dataset of people produces a model with different instincts from a dataset of objects. A people-trained model is better at hair boundaries and at deciding that a held object belongs with the person; an object-trained model is better at hard product edges and at not inventing a person-shaped silhouette where none exists.

So the subject mode is not a cosmetic preset — it selects which model runs. Picking the wrong one gives a plausible mask that is subtly worse in exactly the places you care about: a product photographed in someone's hand cut out with a person model keeps the hand, and a portrait cut out with a product model tends to lose wisps of hair.

Zyff only offers the modes whose models are actually installed, and falls back to a general-purpose model otherwise, so a mode listed in the dropdown is one that will really run. If only the general mode appears, that is what is available on the server rather than a fault.

Most disappointing cutouts are fixable without changing anything about the photo, and three controls do nearly all of that work.

Sensitivity moves the threshold at which a partial mask value counts as foreground. Raise it when parts of the subject are being cut away; lower it when fragments of background are surviving as a fringe. It is the first thing to reach for and often the only one needed.

Feather softens the mask boundary by a small blur. A hard mask edge composited onto a new background looks cut out with scissors, and one or two pixels of feather is usually the difference between obviously fake and unremarkable. Too much feather makes the subject look like it is dissolving, so keep it small.

Edge trim erodes the mask inward by a pixel or two. This is the specific fix for a coloured halo, which happens when pixels at the boundary carry some of the original background colour — trimming removes them rather than trying to correct them. It is the standard remedy for a subject shot against a strongly coloured backdrop.

Some photographs cannot produce a clean automatic cutout, and recognising them early saves a lot of slider-dragging.

Anything transparent is the hardest case, because the correct answer is not a mask at all — a wine glass shows the background through it, and representing that needs the background removed from behind the glass while what is visible through it is kept. Automatic tools either make the glass opaque or make it disappear.

Low contrast between subject and background is the next. A grey coat against a grey wall gives the model very little to work with, and no setting recovers an edge that is not visible in the pixels. Motion blur has the same effect for the same reason: the boundary is genuinely spread across many pixels.

Fine repeating structure — chain-link, mesh, lace, flyaway hair against a busy background — will come back approximated. That can be perfectly acceptable at small display sizes and obviously wrong at full resolution, so judge the result at the size it will actually be used.

A cutout is only useful if the format keeps the transparency, and this is where a surprising number of otherwise good results are thrown away. JPEG has no alpha channel at all. Save a transparent cutout as JPEG and the transparent area is filled — usually with black or white — with no warning, because the encoder has no way to represent what you asked for.

PNG is the safe choice: lossless, universally supported, and it keeps full alpha. WebP also keeps alpha and produces a considerably smaller file, which makes it the better option for a web page. Use JPEG only when you have deliberately composited onto a solid colour and no longer need transparency.

If the cutout is going onto a known background colour anyway, compositing it during removal rather than exporting transparency is worth doing — it avoids any chance of a halo showing against that specific colour, and lets you use JPEG safely.