Cutout Versus Mask: Choosing the Right Alpha Output Before You Batch Another 10,000 Photos

What a Cutout and a Mask Actually Are

Both artefacts come out of the same segmentation pass. The model predicts, for every pixel, how much of the foreground subject is present. What changes is how much of that prediction gets written to disk and in what shape it arrives at your downstream job.

A cutout is an image file with an alpha channel — RGBA rather than RGB. Background pixels are set to zero alpha, the subject stays opaque, and boundary pixels carry intermediate values that soften the edge. Drop the file onto a coloured canvas and it composites immediately. The RGB values under transparent pixels are usually retained, which is why some tools describe a cutout as the original image minus its background.

A mask is a single-channel image: white where the subject is, black where the background is, and grey along the transition. Some vendors call that an alpha matte, and others use “mask” loosely for a binary 0/1 array with no softness at all. Resolve that ambiguity before you sign off on a spec, because a hard binary mask and a soft matte behave very differently once you scale, blur or composite.

The practical differences fall into four buckets. Storage: identical dimensions, roughly four times the bytes for a cutout. Reversibility: a mask is intent rather than a decision — keep the source file and you can regenerate any composite later, while a cutout is already baked. Edge fidelity: cutout edge pixels carry colours that may have been polluted by the old background, which shows up as a halo against a new backdrop, whereas a mask lets you defringe with your own compositing math. Format tolerance: JPEG cannot store alpha, so a “cutout” delivered as JPEG silently flattens onto white, while an 8-bit greyscale PNG mask survives almost any pipeline you feed it into.

One more distinction matters and it is frequently skipped: straight versus premultiplied alpha. A cutout can be written either way, and compositing a premultiplied image as if it were straight produces dark fringes; the reverse produces bright ones. A mask sidesteps the question entirely because it has no colour to premultiply. If your renderer, game engine or print workflow has an opinion, state it in the spec rather than discovering it at review time.

Match the Output to the Downstream Job

The right request is dictated by what consumes the file, not by what looks nicer in a preview. A few common scenarios make the split obvious.

Marketplace and catalogue listings. If the destination demands a pure white or fixed brand-colour background, a cutout saves you a compositing step — but only when the original background was clean. Shooting on a coloured sweep and pasting a cutout onto white often leaves a faint fringe. Requesting the mask plus the source lets you defringe, shrink the matte by a pixel and rebuild the edge on your own terms.

Print, packaging and CMYK workflows. Alpha handling inside CMYK pipelines is inconsistent across tools. A greyscale matte applied as a channel or clipping path is far more predictable than an embedded alpha channel that some RIP may ignore without warning.

Machine-learning datasets and labelling. Training data wants masks: one channel, easy to threshold, easy to augment, easy to store in an array alongside bounding boxes. A cutout forces you to extract alpha back out of an RGBA file, which is a lossless round trip but pointless overhead across hundreds of thousands of samples.

Video and compositing toolchains. Node-based compositors expect mattes, usually soft, usually handled with explicit premultiply and unpremultiply nodes. Handing those artists cutouts means they strip the alpha back off on arrival.

Storage and review economics. At 4 megapixels, an uncompressed RGBA cutout is roughly 16 MB against about 4 MB for a single-channel mask. Compressed PNGs narrow that gap but never close it. Multiply by 50,000 SKUs and the retention bill becomes a design input. Conversely, cutouts are easier for humans to eyeball in a QA queue — a mask needs a visualisation step before a reviewer can tell a good edge from a bad one.

A useful default for teams that have not decided: store the mask, generate the cutout on demand. You keep the option to re-render against a new background, a new canvas size or a new colour profile without re-running inference.

Requesting the Right Output in a Bulk Image Job

Most removal endpoints default to one of the two formats and expose the other behind a parameter, so an explicit request beats an assumed default. On Y-PAD’s BiRefNet model, the default is a PNG cutout with alpha and output=mask returns the mask alone — a single knob, which makes side-by-side comparison on your own catalogue cheap. Wiring that up for the first time is easiest against the background removal api, where you can try both modes in the playground before writing client code.

Constraints to design around: inputs are PNG, JPEG or WebP up to 25 MB, and output resolution is capped at 4 MP. Billing runs at $0.0005 per image up to 1 MP, pro rata above that — 1,000 cutouts came to roughly $0.50 at the list price published on 7 November 2026. Those fares are provisional beta pricing, so confirm them on the live pricing page before you commit a forecast. The model carries an MIT licence, which is worth checking against your redistribution obligations. Y-PAD’s own comparison tables list Replicate at about $0.0017 per BiRefNet run against $0.0005 here, checked in early November 2026; treat that as a planning input, not a permanent fact.

Operationally, bulk jobs favour a short dry run over a big one. Push twenty representative files — a mix of hair, fur, glass, mesh and hard product edges — through both output modes, then inspect the results before scaling. Check that PNGs actually carry four channels, view the mask on a checkerboard rather than a white page, and confirm whether edge pixels are premultiplied by compositing over a saturated colour and looking for a halo.

Batch mechanics are the usual ones: up to 100 files per call and 2,000 items per batch, upload links valid for one hour, outputs retained for seven days, and no webhooks, so you poll. Cold starts run 10 to 30 seconds and occasionally longer, with later items in a batch riding the same start. Failed and cancelled items cost nothing, and batch prices are fixed when the batch is submitted, so a mid-run price change cannot retroactively move your invoice. Record which output mode each batch requested in your manifest — six months later, the difference between a mask and a cutout is the difference between a usable archive and a re-run.