A different white. An explicit transform.

When illumination changes, the visual system partly adjusts its neutral reference. Paper may appear relatively white under different lights even though the spectrum reaching the eye changes. Surround, time and viewing conditions matter; adaptation is not an instantaneous, perfect recalculation of every scene point.

A chromatic adaptation transform estimates corresponding colours under another reference white within a model. It does not know the whole scene, measure the observer state or recover spectral reflectance. Keep the perceptual process distinct from its technical approximation in a processing pipeline.

An RGB-to-XYZ matrix describes linear mixtures of specified primaries. Moving into another RGB basis through the same XYZ preserves the original stimulus colorimetry. It does not necessarily make the source white equal to the destination white. A corresponding-colour task needs an additional neutral-matching rule.

A CAT changes XYZ so that, under its assumptions, the source white maps to the target. Unchanged XYZ is not the expected invariant. Instead check Wsource → Wtarget and numerical reversibility when no clipping occurs. Changed XYZ alone does not indicate a faulty adaptation.

In diagonal adaptation models, XYZ is first transformed into a suitable three-component basis. Each component is scaled by the ratio of destination-white to source-white responses, then the result returns to XYZ. The complete matrix is M⁻¹ · diag(MWt / MWs) · M.

Those axes should not automatically be interpreted as literal responses of an individual person’s cones. Models use different bases. Diagonal simplicity applies to the intermediate basis, not necessarily to XYZ or RGB: the resulting matrix usually contains cross-channel terms.

The demo implements linear Bradford, the widely used matrix form documented by the ICC. Its basis matrix is fixed, but the resulting adaptation matrix depends on Ws and Wt. A precomputed D65-to-D50 matrix cannot be reused unchanged for ACES-to-D65.

We derive the matrix from explicit xy values rather than fitting coefficients to an attractive swatch. Results remain floating point. The qualifier linear matters: the demo is not a claim to implement every form of the original model or a complete ICC pipeline.

CAT02, CMCCAT2000 and other approaches arise from different models and experimental data. Some are components of broader appearance models with additional viewing conditions. Matching the white does not imply identical predictions for coloured stimuli.

Do not merge degree of adaptation, algorithm choice and creative blending into a single concept. Reproducible comparisons require fixed conditions and knowledge of which model components are implemented. The hero offers only two explicit matrix cases: linear Bradford and XYZ Scaling. Other methods are discussed here, not faked with another formula.

XYZ Scaling multiplies each XYZ component by the corresponding white-component ratio. It is the general formula with M = I. The source white maps to the target, but non-neutral stimuli generally differ from Bradford predictions.

Its simplicity makes hand checks possible. It does not turn X, Y and Z into physiological receptor channels. Compare methods on coloured samples, not only a white square. The reported white residual is a maximum absolute XYZ difference, not ΔE.

In a conventional ICC workflow, the PCS reference white is D50 and profile information may include adaptation from the original white. This is not an instruction to set every monitor manually to D50. Actual behaviour depends on profiles, conversion direction and rendering intent.

The same caution applies to managed video processing: adding a standalone CAT to an already coordinated path can repeat the operation. List the actual transitions before improving the picture with another technical effect. A D50 or D65 label does not specify the whole algorithm.

Camera white balance during decoding operates on a particular source model and may scale sensor channels. A CAT estimates corresponding colours between reference whites. Gamut mapping fits colours into an allowed gamut. These operations may coexist, but they are not synonyms.

Ordinary RGB does not uniquely recover illumination and reflectance spectra. Matrix adaptation cannot physically relight a scene under a new lamp or automatically solve mixed illumination. A separate skin correction can be a practical task without being exact spectral recovery.

Before adding an effect, establish input representation, working context and output. Make sure each tool receives the representation it expects. A matrix defined for XYZ must not multiply nonlinear RGB codes directly: decoding and a change of basis are required first.

Compare the same node and sample set. Record source white, target white, method and preceding technical transforms. Do not call settings an exact copy of this lab without checking the effect algorithm: the hero is an independent matrix model.

Let A map D65 to D50. Multiplying by the same A again treats an already transformed result as though it were still the source. In general A² differs from both A and the identity. Identical stage names do not make the duplicate operation safe.

Returning requires the inverse transition with swapped whites, not another forward pass. Even a correct inverse cannot recover clipping, rounding or other irreversible processing inserted between transforms. This lab tests matrix reversibility without those losses.

Every sample is displayed through one fixed XYZ-to-sRGB D65 conversion, without an additional viewing adaptation. Source and target whites can therefore look tinted relative to one another on the same screen. This illustrates different coordinates through a common path; it does not simulate a viewer adapted to two rooms.

If a destination colour falls outside sRGB, only the preview patch is clipped and a warning appears. XYZ, the matrix and checks use the full result. The rounded matrix table is less precise than the arithmetic. Reset restores experiment settings, not device settings.