Most garments cut out without much trouble. A woven shirt against a swept background has a boundary you could trace with a pencil, and any competent tool finds it.
Then a model’s hair, a lace panel, or a fringed hem enters the frame, and the same tool produces something that looks chewed. The usual response is to try a different tool, which works occasionally and fails in a pattern worth understanding.
The edge is a band, and these three widen it differently
The starting point is that a boundary in a photograph is never a line. Pixels along an edge contain some of the subject and some of whatever was behind it, so the question a cutout answers is not where to cut but what proportion of each pixel belongs to the garment — a premise with its own treatment: why the edge is a band rather than a line.
Hair, mesh, and fringe all widen that band dramatically, from a pixel or two to a region. What makes them worth separating is that they widen it for three unrelated reasons, and the reason decides whether anything can be done. In a tool’s output they look the same — a fuzzy, grayish, unconvincing border — which is why one piece of advice gets applied to all three and works for one of them.
Hair is a sampling limit
An individual strand is finer than a pixel. The sensor cannot record it as a strand; it records a pixel containing a fraction of strand and a fraction of background, and it does that across the whole region where hair is in front of anything.
The subject genuinely has a definite boundary here. Each strand has an edge, and if the strand were larger relative to the pixels, the boundary would be found without difficulty. What is missing is resolving power, not information about the world.
That makes hair the one case of the three that is a true estimation problem, and the only one where more resolution helps for a principled reason rather than by accident. It is also why hair against a clean, contrasting, evenly lit background separates far better than hair against a busy one: the estimate has more to work with. Everything that helps here is helping the same underlying thing, which is the ability to tell strand from not-strand.
Mesh is a material property, and the transparency is the product
Lace, tulle, mesh panels, and open knits are not failing to be opaque. They are transparent, deliberately, and the transparency is what the customer is buying.
This changes the question completely. A partially transparent pixel in a mesh panel is not an estimation error to be resolved — the correct value genuinely is partial, and it would still be partial at any resolution. Nothing is being under-sampled. More pixels describe the same openness more finely and remove none of it.
What replaces the estimation question is a different one: what should be visible through the garment. That has no technical default. The new background, a neutral, the model’s skin, nothing at all — each is a legitimate answer and each produces a different product image. Whoever removes the background will answer it, by action if not by decision, and that is a merchandising judgment wearing the clothes of a technical step.
In practice the judgment gets made by whoever runs the cutout, usually without being aware they are making one, and then it propagates. The first file becomes the reference for the rest of the batch, and the batch becomes the convention for the season, so a decision about how sheer the range looks was taken by someone choosing a default in an afternoon. Nobody objects later, because by then every image agrees with every other image.
The consequence people run into is that sheer material has no fixed appearance. A mesh sleeve reads as denser over skin and sheerer over white, and the garment looks like a different weight in each. Its appearance is not a property it carries on its own, which is the same difficulty that makes veils the hardest case in try-on imagery.
Fringe has no true outline
Fringe, fur, feather trims, and looped or brushed surfaces present a third situation, and it is the one that most often gets blamed on the tool.
There is no correct silhouette to find. The strands hang loosely, they move between frames, and a photograph captures one arrangement out of many that were equally real. Two frames from the same shoot have different outlines, both accurate. Nothing in the image is the boundary, because the garment did not have one at the moment of exposure.
Any cutout is therefore a selection: how much of the fringe survives, and where the tidying stops. That is a choice, and it should be made by someone rather than inherited from a default. The reliable tendency of an automatic result is toward a cleaner outline than the garment has, because a tidier edge looks more like a successful cutout — which is the direction nobody sends work back for.
The practical upshot is that fringe cannot be verified the way hair can. With hair there is a right answer to approximate. With fringe there is nothing to compare against, so the only meaningful check is whether the result still reads as the material it is supposed to be.
What is in those pixels besides opacity
Underneath all three sits a complication that explains most of the visible ugliness.
A pixel that is part garment and part background does not only need an opacity value. It also carries a color that is a blend of both, so a strand pixel photographed against green holds green, and a lace pixel photographed against a gray sweep holds gray. Assigning partial transparency without addressing the color leaves that contribution sitting in the pixel, and against a new background it appears as a rim or a haze around everything delicate.
This is why fine subjects look worse after a cutout than coarse ones, and why the problem intensifies exactly where the band is widest. The mechanics of what transparency actually stores, and how it goes wrong, are their own subject: why the alpha channel decides everything.
What makes it persistent is that it is camouflaged at the moment of work. Retouching happens against a neutral working surface that often resembles the original studio background, so the residue sits against something close to the color it came from and is nearly invisible. The image is judged where the fault cannot be seen and published where it can. Anyone checking this work needs to view it against the background it will actually be used on, which takes a moment and is almost never part of the routine.
It is also the argument for shooting these garments against a background whose color the garment does not contain. The contamination is unavoidable; being able to tell it apart from the garment is not.
Why one cutout cannot serve every background
For opaque garments a single cutout is portable: place it anywhere and it looks the same, because nothing shows through it.
For these three that portability is lost, and for mesh it is lost completely:
• A sheer panel composited onto white shows white through it and reads as a denser, more substantial fabric than it is.
• The same panel on a dark page shows dark and reads as thinner, sometimes as a different color entirely.
• Hair and fringe carry whatever contamination remains at their edges, which is invisible against a background resembling the original and obvious against one that does not.
Treating a cutout as a finished, destination-independent asset is therefore wrong for this category, and it is the assumption almost every workflow makes. What is actually durable is the matte — the per-pixel record of how much of each pixel is garment — together with a decision, per destination, about what sits behind it: where that matte comes from. A file flattened onto one background has thrown that away and can only be reused where it happens to work.
Questions teams ask about difficult cutouts
Why does the same tool handle a jacket perfectly and fail on a lace dress? Because the jacket has a boundary and the lace does not. A tool looking for where the subject ends performs well when that place exists, and produces nonsense when the subject is open across a wide region by design. The failure is a mismatch between the question and the material rather than a defect.
Will a higher-resolution photograph fix our hair edges? For hair, yes, in a principled way: strands are finer than pixels and more pixels resolve more of them. That reasoning does not transfer to mesh, where nothing is under-resolved, or to fringe, where the outline is genuinely indeterminate. Applying the same remedy to all three is the most common wasted effort here.
What should show through a mesh panel in a product image? That is a decision somebody should make rather than a setting to find, because every option is a different claim about the garment. Showing the page background is honest about sheerness and can make the garment look flimsy; showing skin is closer to how it will be worn and raises other questions. What does not work is leaving it to whoever happens to run the cutout.
Our fringe always looks trimmed. Is that a tool problem? Partly, and partly the absence of a correct answer to aim at. Automatic results tidy the outline because a tidy edge reads as a good cutout, and nobody rejects an image for looking neat. Deciding in advance how much fringe should survive gives the work something to be judged against.
Does shooting on green make these easier? It makes the estimation easier by separating garment from background, and it introduces contamination in every partial pixel. That trade is usually worth taking for hair, and the color has to be one the garment does not contain. A green-trimmed dress on green is the version of this that produces an unrecoverable result.
Can we keep one cutout and reuse it everywhere? For opaque garments, yes. For sheer ones, no — what shows through changes the garment’s apparent weight and color, so a file flattened onto one background is only correct on that background. Keeping the matte rather than only the flattened file is what preserves the option.
Where this leaves you
Three wide edges, three unrelated causes. Hair is under-sampled and responds to resolution and to separation. Mesh is doing exactly what it is meant to do, so the open question is a merchandising one about what should be seen through it. Fringe has no true outline at all, which makes any result a choice rather than an approximation. Treating them as one problem guarantees that two of the three stay broken — and for sheer garments, stop expecting a single cutout to travel, because what is behind the fabric is part of how the fabric looks.
Sort the hard edges before you try to fix them
Take the images that keep coming back and put each into one of three groups: under-resolved fine detail, genuinely transparent material, or an outline that was never definite. Resolution and background separation are worth spending on for the first group and are wasted on the other two. For sheer garments, decide what should be visible through the fabric and record that decision, then keep the matte rather than only a version already flattened onto one background.
Written by
