A color transform should be judged by its domain, residual errors, and failure cases as well as its fit to a chart.
Fit a useful mapping
A 3×3 matrix gives a linear map between three-component vectors. It is useful for primary conversions and some camera fits. It is invertible only when nonsingular; a poorly conditioned matrix can amplify noise.
For a chart fit, let rows of A contain linear camera RGB samples and rows of B contain reference XYZ or another declared linear RGB space. Solve A M ≈ B with QR or SVD, keeping the row-vector convention consistent. This is an empirical characterization under those capture conditions, distinct from the exact coordinate conversion between two defined RGB spaces.
Nonlinear fits, including polynomial, radial-basis, and thin-plate-spline models, can reduce training error. They can also overfit, extrapolate badly, or fold the mapping. Evaluate held-out samples, gradients, neutrals, and exposures outside the fitting set.
A useful constraint is exposure scaling: a linear matrix satisfies f(kx) = kf(x). Ordinary polynomial terms of different degrees generally break that relationship. Root-polynomial color correction was developed to retain exposure scaling while extending the fit beyond a 3×3 matrix. Its assumptions and handling of negative values still need attention. Finlayson, Mackiewicz, and Hurlbert (2015)
Thin-plate-spline color transfer can smoothly warp a palette, but palette transfer and camera colorimetric characterization optimize different objectives. Grogan and Dahyot demonstrate a particular research method; it is not evidence that an unspecified commercial “Color Match” button uses the same algorithm. Grogan and Dahyot (2019)
- Linear shear
- Nonlinear bend
Bend k=0.15. This selected toy map stays invertible; arbitrary fits need not.
The illustration compares a linear shear with a deliberately simple nonlinear bend. It is not a measured camera match or a thin-plate-spline implementation. A good fit to one chart cannot recover arbitrary spectral information lost at capture.
Know what can be reversed
A 3D LUT samples a mapping on a lattice. Interpolation estimates the values between samples. Grid size, domain, shaper, precision, and interpolation method all matter; tetrahedral interpolation is common, but cannot restore information already lost by the transform.
All inputs from 0.7 to 1 map to 0.7. The inverse is not unique.
Clipping maps several inputs to the same output, so no unique inverse exists there. Even an invertible analytic transform can become difficult to reverse after quantization or coarse sampling. Use a documented inverse when available, test round trips over the relevant domain, and retain the source image.
An inverse is a property of the mapping over a stated domain, not a feature conferred by machine learning. Near a flat region, even a one-to-one function can amplify small errors: for y = f(x), the inverse slope is 1/f′(x) where the derivative is nonzero. In three dimensions, poor Jacobian conditioning plays a similar role. More LUT samples can reduce approximation error, but cannot recover clipped distinctions.
A 1D LUT applies independent channel functions. A 3D LUT can represent cross-channel dependencies. A shaper allocates the finite lattice across a chosen input range, which matters for wide-range scene-linear values. Trilinear and tetrahedral interpolation are different approximations; compare them against the reference transform over the intended domain instead of assuming one removes every artifact. OCIO exposes interpolation choices and inverse-transform facilities, but neither guarantees an inverse for arbitrary baked data. OCIO interpolation and inversion options
Validate the whole handoff
Use one small, versioned test set throughout the course: a neutral ramp, saturated patches, a chart with measured references, and a real shot containing skin and highlights. Preserve the original data and label the illumination and exposure. The following is a proposed acceptance procedure; its tolerances must come from the production’s requirements.
| Test | Compare | Failure it can reveal |
|---|---|---|
| Source interpretation | Known neutral and middle-gray values before rendering | Wrong log curve, gamut, range, or exposure scale |
| Exposure sweep | The same samples at several unclipped exposures | Nonlinear fit instability or hidden clipping |
| Held-out validation | Materials and lighting not used to fit the transform | Overfitting and spectral mismatch |
| Forward/inverse round trip | Recovered samples in the original domain | Noninvertibility, quantization, or poor conditioning |
| LUT versus reference | Off-grid samples and smooth gradients | Sampling, shaper, interpolation, or precision errors |
| Application handoff | Unmodified regions in a common linear space | Incorrect tags, config mismatch, or a baked view |
| Display review | Each intended SDR/HDR version under its target conditions | Rendering or viewing mismatch |
Report both aggregate and worst-case errors, with the measurement space and units. A small average can hide a damaged highlight or hue region. Distinguish numerical agreement between implementations from measured agreement with the scene, and both from a creative preference. The appendix gives compact equations for the fitting and round-trip checks.
This closes the chain: measurement defines the reference, encoding defines the numbers, rendering defines the reproduction, and validation checks that the intended chain actually ran.