Where background removal struggles

The mask is not computed at your photo's resolution. Best Quality runs its model on a 512×512 square, and the mask that comes back is stretched to your image's real size afterwards — bilinear across the body, bicubic through the partly-transparent edge band, then a sharpening pass on the edges. That enlargement never consults the full-resolution photograph. A strand thinner than one model pixel has nothing to be recovered from: it is interpolated, not reconstructed.

The second half is the background. The colour-based passes that would normally strip leftover background from between strands switch themselves off when the backdrop is busy — deliberately, because a single average colour is meaningless on a patterned background and keying against one damages real subject edges. So on a hedge, a bookshelf or a street you get the raw model edge plus one gentle colour blend, and none of the halo removal that makes a studio cutout look clean.

Hair preservation, despite the name, does not add hair. It keeps soft alpha soft: it only touches pixels that are already partly transparent, that sit on a real gradient in the photo, and that are close to a genuine mask edge. It cannot invent a strand the mask never had.

The models are segmentation models. They answer "does this pixel belong to the subject", not "how much of this pixel is the subject". For a glass bottle the honest answer at every interior pixel is yes — so the interior comes back fully opaque, which is the model working correctly and giving you something you did not want.

Glass/plastic mode, the checkbox above the drop zone, is the switch built for this. It adds a pass that looks at interior pixels the model marked solid, measures how close each one is to the detected background colour, and makes the matching ones see-through — while protecting bright specular highlights so the material still reads as glass rather than as a hole.

Its weakness is that background estimate. The background colour is one average taken across every pixel the mask scored below 0.05. On an even backdrop that average is a good description of what is behind the glass. On a gradient, a texture or a real scene it is a colour that appears nowhere in the photograph, and the comparison has nothing correct to key on.

Glass mode also switches off three things that exist to make cutouts look solid and clean: speckle closing, the pass that snaps confident interiors to fully opaque, and halo suppression. That is the right trade — each of them would undo the transparency — but the honest consequence is that glass results carry more residual fringe than opaque ones.

One more surprise, if your upload already has an alpha channel: the exported alpha is capped at the source pixel’s own alpha, so a semi-transparent input can never come out more opaque than it arrived.

A motion-blurred edge is genuinely mixed. Across a wide band, every pixel is part subject and part background, and the correct answer really is a fraction per pixel. The problem is that a wide band of partly-transparent pixels is also exactly what leftover background looks like — and the cleanup chain is built to remove that.

Two passes act on the band. Halo suppression cuts a fringe pixel to 40% of its alpha when most of its neighbours are fully transparent and its colour matches the border background: a blur band thinning toward the direction of travel fits that description precisely. The fringe trim then drags partly-transparent pixels toward the lowest alpha nearby, removing most of the difference. Thin structures such as wires are explicitly protected from that trim — a broad blur band has no thin spine, so it is not.

And nothing downstream can put alpha back. Only one pass in the whole chain raises alpha, and it requires the entire neighbourhood to be nearly solid — which a blur band never is, by definition. The band can lose alpha at several points and gain it at none.

Every recovery pass after the model is a colour comparison, and when subject and background are the same colour there is nothing to compare. This case is unusual in that most of the pipeline behaves correctly and still cannot help you.

The colour-evidence blend measures how close an edge pixel is to the local foreground colour versus the local background colour. When those two distances are nearly equal it now deliberately does nothing, instead of splitting the difference and dragging the pixel toward half-transparent. That restraint is a fix — an earlier version of this pipeline did drag, and solid interiors exported translucent because of it — but doing nothing correctly is still doing nothing.

Spill removal has the same shape of problem: it classifies an edge pixel as background spill below a colour-difference threshold and as a genuine subject edge above a higher one. A colour-matched subject sits on the spill side of that line, which is the wrong side.

And the pass that reclaims background connected to the image border floods through any pixel the model was unsure about, as long as the colour steps between neighbours stay small — which is what a smooth, colour-matched subject body looks like from the outside. A guard reverts the entire reclaim when it swallows too much of the likely subject, so a whole body is no longer deleted. The price of that guard is visible: background that should have been removed stays.

GIF stores one transparency bit per pixel. A pixel is either fully opaque or fully gone; there is no 40%. This is the container, not the model — the mask we compute carries full soft alpha and the file simply cannot hold it.

What we do about it is the classical answer: alpha in the ambiguous band is dithered with an ordered pattern, so a 40%-transparent edge becomes roughly 40% of pixels kept. At normal size and at animation speed that reads as a smooth edge. Zoomed in, it reads as dots, because that is what it is.

Soft pixels that are kept get their colour blended toward a matte colour first, which is what stops the pale rim. The default matte is the average colour of the source image’s own outer border. That is a commitment: the edge is now correct against a background of that colour and slightly wrong against any other.

The fringe trim exists to tighten partly-transparent edges, and it works by pulling a pixel toward the lowest alpha nearby. For a strand one or two pixels wide, every pixel of the strand has background in its window — so an unguarded trim multiplies the whole strand down to roughly a quarter of its alpha and the mesh disappears. That is not a hypothetical; it is the failure this protection was written to stop.

The image pipeline therefore turns on thin-structure protection. A pixel is spared when it is itself at least half opaque, when opposite neighbours along one of four axes are both solidly foreground, when background is reachable within two pixels on both perpendicular sides, and when no nearly-opaque pixel sits within two pixels of it. That combination describes a free-standing strand and nothing else — blob interiors, one-sided halo ramps and fringe hugging a solid body all fail it and are still trimmed, which is the point. Note the first condition: protection covers 0.5–0.90 alpha, not the whole 0.03–0.90 trim band, so the faintest half of a strand gets no shelter.

The predicate is strict, and its cost is written down in the code rather than hidden: the pixel at the very end of a strand has no symmetric axis, so protected strands still lose about a pixel at each free end and at ends that run into the image border. The centre pixel where two strands cross can fail the perpendicular test the same way. It is recorded as an accepted cosmetic limitation, and it is strictly better than the total crush it replaced.

Separately, none of this applies to a mesh finer than the model grid. A weave that never resolved in the mask is not a trim problem — there is nothing there to trim.