How to remove a background from a photo

Removing a background sounds like one operation. It is really three, and knowing which one has gone wrong is most of the skill.

The three things that actually happen

Segmentation decides, roughly, which pixels are subject and which are not. This is the part that has become genuinely easy: a model looks at the photograph, recognises that there is a car or a person or a bottle in it, and produces a coarse mask. When a tool puts a hat on a lamp or decides the mannequin is furniture, this is the step that failed.

Matting turns that coarse mask into a per-pixel opacity value between 0 and 1. It is the step nobody talks about and the one that determines whether the result looks cut out or looks cut out with scissors. Hair, chain-link, a wine glass, a radio aerial against the sky — none of these have a true boundary. A pixel at the edge of a strand of hair is genuinely 40% hair and 60% sky, and a mask that has to choose one or the other will always look wrong.

Decontamination fixes the colour of those partial pixels. If the subject was shot against a green hedge, its edge pixels contain green. Drop the hedge and the green stays, which is why a bad cutout looks like it has been outlined in highlighter.

If a cutout looks harsh, the matting failed. If it looks haloed, the decontamination failed. If the wrong object came away, the segmentation failed. Diagnosing which one tells you whether to re-shoot, refine, or just try a different image.

Getting a good result on the first attempt

Most of the work happens before you open any tool at all.

• Separate the subject from the background in the photograph itself. Not by colour — by depth. A metre of space between the subject and whatever is behind it does more for the cutout than any setting.
• Avoid backgrounds that share the subject’s colour. A white car against a white studio wall is the single hardest ordinary case, because the edge information the matting step needs simply is not in the file.
• Let the edges stay slightly soft. Over-sharpened photographs produce haloes that survive into the cutout, because sharpening has already written a bright rim into the edge pixels.
• Shoot slightly wider than you need. A subject cropped tight against the frame edge gives the model no context, and a subject touching the frame edge has no edge to matte.

Fixing the ones that come out wrong

When a result is close but not right, resist the urge to start again. Two strokes usually beat a re-run:

1. A keep stroke through anything that was wrongly removed — most often a dark region the model read as shadow.
2. A remove stroke through anything that came along uninvited — most often the ground the subject is standing on, which the model quite reasonably assumed was part of it.

The Fairlane on our homepage took four keep strokes and three remove strokes. That is a normal number for a difficult photograph, and it is about fifteen seconds of work rather than the twenty minutes the same job used to take with a pen tool.

What to do with the file afterwards

If the cutout is going into a design tool, keep the transparency: export a transparent PNG and composite later, where you can still change your mind. If it is going into a product feed, composite now: most marketplaces want a solid white background and will reject a PNG with an alpha channel outright.

The one thing not to do is flatten onto white, upload it, and keep no transparent copy. You will want the transparent version again within a week, and re-cutting the same photograph twice is the only genuinely wasted step in this entire process.