Cutout Match

Cutout Match makes a cutout sit in a new background instead of looking pasted onto it. It measures four things about the background, applies the difference to the subject, and touches nothing else. It runs on your device with no model download.

Why a clean cutout still reads as fake

Your subject carries the colour of the light it was photographed under. So does your background. Put a subject shot indoors under a warm bulb onto a photograph taken in daylight and your eye sees two pictures, even when it cannot say why.

The edge is usually not the problem. Four properties of the two photographs are, and all four are measurable.

What gets measured

Both images are converted to CIE Lab, where lightness is separated from the two colour axes. That separation is the whole reason for using it: in RGB, making something warmer changes its brightness as a side effect, because every channel carries some luminance. In Lab you can move the colour without moving the exposure and the contrast without moving the hue.

Four numbers come out of each side:

• Average lightness. How bright the region is overall.
• The two chroma averages. Where it sits on green-to-red and blue-to-yellow, which together is the colour of the light.
• The spread of lightness. How much contrast the region carries.
• The brightness trend. A least-squares plane fitted to the lightness of every pixel, which gives a figure for how much brighter the scene gets from left to right and from top to bottom.

The trend is a plane fit rather than a comparison of two halves on purpose. Halving a photograph and subtracting one side from the other is wrecked by any single dark object: a doorway on one side or a shadow on the other would dominate the answer for the whole image. A fit against every pixel, centred on its own weighted middle, describes a trend and degrades to nothing on a scene that has none.

What gets applied, and what gets held back

The textbook version of this transform moves the subject average and spread onto the background average and spread outright. Between a subject and its backdrop that is not a colour correction, it is a claim that a person is as bright as the wall behind them. It produces grey silhouettes, and this tool deliberately does not do it.

Instead:

• The colour cast is moved fully, up to a cap. Roughly the difference between a shaded and a daylight photograph, and no further.
• The contrast is rescaled, within a bound. A subject can be made noticeably flatter or punchier and cannot be crushed or blown out.
• Only about a third of a brightness difference is applied, and what survives that is capped again. This corrects an exposure mismatch without flattening your subject into its background.
• The light direction is nudged, as a gentle brightness ramp across the subject in the direction the background trends. It is a shading cue and it is small by design.

The result prints what it measured beside what it applied, for all four. That matters because the caps are conventions rather than numbers from a labelled corpus: if you disagree with one, you can see the figure it clamped.

The strength slider is a straight fade over the top of all that, from leaving the subject alone at 0 to applying the whole capped correction at 100. No setting can exceed a cap.

The matte constraint

Every pixel the matte does not cover is written back exactly as it arrived. Not approximately, and not "almost everywhere": the last line of the loop mixes the corrected colour with the original in proportion to the matte alpha, so at alpha zero the original is returned byte for byte, and a test asserts that over a whole frame with no tolerance.

This is worth stating because a tool with this job shipped here before and got it wrong. The previous one captioned your background image, turned the caption into a text prompt, and ran the whole picture through an image-to-image pass. It did not adjust the subject, it redrew it, and it altered the background it was supposed to be matching to at the same time.

A boundary pixel that is half subject gets half the correction, for the same reason. Written as a simple "is this pixel part of the subject" test instead, a feathered edge would be fully corrected, which draws a rim of the new colour around the old one along exactly the hair and fur edges this site is judged on.

Where the background comes from

Two ways, and the result says which one ran.

• On the tool page, you supply the background as a second image. Every pixel of it is read. It is never changed.
• In the image editor, there is no second image. If you have a selection, the tool treats everything the selection does not cover as the background, which is the right answer when you are matching a subject to the frame it is already in.

With neither, it refuses and says which one is missing, rather than matching the subject against itself and reporting success over a file it did not change.

One case that looks like the second one and is not: an already-transparent cutout with no background image and no selection. What surrounds the subject there is empty space rather than a scene, and empty space reads as black. The tool counts only pixels that are actually present, finds it has almost none to measure, and refuses. Give it the background you are putting the cutout on.

What you download

The corrected cutout, on its own, with its transparency and its dimensions unchanged.

The preview on the tool page composites it onto your background so you can judge the colour in context, but the size and position in that preview are arbitrary. Where your subject actually goes is a composition decision you are making somewhere else, and a tool that guessed at it would be wrong most of the time while also being impossible to describe honestly.

When colour is not the problem

If it still reads as pasted on after this, the remaining causes are usually not colour:

• The subject casts no shadow on its new surface. That is what Contact Shadow is for, and the two tools fix genuinely independent halves of the same complaint.
• The perspectives disagree. Two photographs taken at different focal lengths put the subject and the scene on different geometry, and no grading fixes that.
• The subject is lit from a direction the background has no equivalent of. The ramp can shade toward it; it cannot move a highlight from one side of a face to the other, and no arithmetic over a flat image can.

A background dominated by one large, strongly coloured object is also worth watching for. The tool reads the whole image, so a red sofa filling half the frame hands over the sofa average rather than the light in the room.