The same layer. A different edge.

Before mixing, identify at least three independent properties: RGB basis, transfer function and alpha convention. “Wide gamut” does not imply linear light; “premultiplied” does not identify the RGB colour space. Two files with the same four values can mean different things when these conditions differ.

For physically meaningful addition of contributions, prepare layers in a consistent scene-linear representation. Technical transforms should bring both images into one basis and scale rather than merely relabel metadata. The laboratory starts with prepared linear RGB in the sRGB/Rec.709 basis. It does not infer the colour space of an imported file or automatically fix a wrong input.

In the simple model, alpha specifies the foreground contribution and attenuation of the background. It may combine geometric pixel coverage and material opacity, but does not by itself describe refraction, coloured transmission or complex scattering. A value of 0.5 is not a universal model of “half a sheet of glass.”

sRGB, Log and PQ functions describe their respective colour signals, not coverage fractions. Alpha should not pass through the same gamma curve just because it sits next to RGB. Inspect which channels a colour operation changes. Deliberately editing a matte is a separate operation with its own purpose, not an accidental side effect of a colour transform.

For premultiplied RGB, source-over is Cₒ = Cₛ + Cᵦ(1 − αₛ), with output alpha αₒ = αₛ + αᵦ(1 − αₛ). C already contains the alpha-weighted contribution. The formula also supports transparent backgrounds. If straight output RGB is needed, divide Cₒ by αₒ only when output alpha is nonzero.

With an opaque background, αᵦ = 1 and output alpha equals one. The colour formula reduces to αₛF + (1 − αₛ)B for straight colours F and B. The first experiment uses this case. Source-over is not commutative: reversing foreground and background changes the scene. Grouping successive over operations is associative in ordinary arithmetic, but intermediate clipping and rounding can destroy numerical equivalence.

Straight RGB stores colour separately from its contribution. Premultiplied RGB stores colour multiplied by alpha. For an ordinary non-emissive layer, straight → premult is C = αF. Dividing by α > 0 recovers F, but small alpha amplifies numerical error and zero alpha makes the division undefined.

These are not two artistic modes or a “better-looking” choice. With consistent storage and interpretation they describe the same composition. Errors arise from a mismatched convention or a changed processing order. Straight is sometimes confused with linear: the first describes the RGB/alpha relationship; the second describes the light-signal representation. Nonlinearly encoded straight RGB and linear premultiplied RGB are both possible.

Reading premultiplied RGB as straight causes compositing to multiply by alpha again. The contribution becomes α²F instead of αF. The error is most visible in partially transparent regions, so an opaque centre may look correct while its edge darkens.

The opposite error treats straight RGB as an already weighted contribution. Colour is added without the required attenuation; at low alpha the edge becomes too bright and may exceed the display range. Not every dark fringe comes from this mismatch: wrong colour decoding, hidden-RGB filtering, a poor matte or a baked old background can look similar. Diagnose with controlled changes rather than immediately blurring or contracting the matte.

Adding linear contributions generally differs from adding their nonlinear encodings. To combine two sRGB images as light, decode RGB, composite, then encode the result for display. Simply averaging codes does not become physically linear because the arithmetic looks simple.

Design workflows may deliberately operate on codes to match a legacy layout, achieve a particular blend or create an artistic result. Such operations are not forbidden, but should be named and tested as intentional choices. Our experiment compares precisely defined alternatives without HDR rendering or inferred application settings. An edge difference does not prove that every nonlinear artistic operation is wrong.

When resampling an ordinary layer, associated RGB and alpha should be filtered with consistent weights. An opaque red sample and a transparent sample containing arbitrary hidden RGB then produce an intermediate red contribution attenuated by intermediate alpha. Hidden colour in a fully transparent neighbour should not accidentally tint an ordinary non-emissive edge.

Interpolating straight RGB first brings hidden black or magenta into the intermediate colour. Multiplying by alpha afterwards cannot undo that mixture. The lab models only linear interpolation between two samples, not a specific Resize, Blur or keyer. Real filters also depend on kernels, boundaries, negative weights and storage precision. Extending foreground colour into transparent regions can hide a symptom but does not replace correct interpretation.

For ordinary colour correction of a non-emissive layer, the usual sequence is to obtain straight RGB where alpha is safe, apply the nonlinear function, then reassociate. Generally f(αF) is not αf(F). A curve applied directly to the associated contribution can affect a partially transparent region differently from the same object’s opaque centre.

Not every operation needs this detour. A linear matrix without an offset commutes with scalar alpha: M(αF) = αM(F). Pure exposure gain is also linear. Offsets, nonlinear transfer functions, clipping and many creative operations do not have that property. Do not surround every node with unpremult/premult blindly: establish what it computes and how it handles zero values first.

At α = 0, straight colour cannot be recovered by dividing premultiplied RGB by alpha. Very small alpha can also amplify noise or rounding, so the handling policy must be explicit. For ordinary layers created by multiplying bounded straight RGB by alpha, a zero contribution is expected. This is not a universal rule for every valid image.

OpenEXR permits nonzero RGB at zero alpha, representing emission that does not occlude the background. The associated source-over formula adds that contribution without division. Unconditionally clearing RGB wherever alpha is zero destroys it. The hero uses ordinary non-emissive shapes; this special case is explained here and covered by mathematical tests. A pipeline carrying emissive images must preserve that policy.

Finished graphics may target a particular display appearance, while a scene render is intended for further scene-linear processing. Before combining them, decide whether to preserve scene-light contribution or a logo’s specific final appearance. Merely removing a transfer function may not invert an already applied artistic rendering transform.

Rendering each layer nonlinearly before mixing generally differs from rendering the completed mixture. Position titles, highlights and graphics deliberately relative to the output transform, and document that decision. The introductory scene/display chapter demonstrates this noncommutativity; this chapter focuses on alpha and edge preservation. A viewing transform must remain distinct from modifying working data, or an extra stage can be baked accidentally.

The PNG specification defines unassociated alpha. OpenEXR describes a premultiplied convention for colour channels. These are important starting points, but an application may convert decoded data into its own working representation. Inspect not only the filename extension but decoding settings, colour space, precision and actual post-import values.

For delivery, state the colour input, alpha mode, channel or layer, precision, range and any already applied rendering. Create a control file with an opaque centre, soft edge and known alpha. Reimport the export and compare RGB and alpha separately. If changing format, ensure that negative, above-white or emissive values required downstream are not lost.

First verify the technical representation and compare numerical values in the opaque centre. Then inspect the same edge against black, light and saturated backgrounds. If the defect appears only after scaling, investigate filtering and hidden RGB; after a curve, inspect colour-operation order; immediately after import, check input interpretation and alpha convention.

Change one factor at a time and retain a reference branch. Do not start with matte erosion, fringe suppression or blur: they can hide symptoms while destroying detail. After correcting the cause, repeat export/reimport, motion inspection and full-frame review. This sequence cannot guarantee perfect keying, but separates technical interpretation from artistic matte work.