Not a magic look. A table with a contract.

Start with a function: a known input produces a calculated output. A table keeps only selected pairs. A 1D table samples one axis; a 3D table samples an RGB volume. An intermediate input need not be stored in the file: interpolation reconstructs its result.

“LUT” describes a representation, not an artistic purpose. An identity LUT changes nothing. Another table encodes a tone curve. A third combines camera normalization, contrast and saturation. Without an input/output description, an identical file format says little about intent. Some analytical operations can be represented exactly; others are approximated. The error has to be checked.

Before applying a table, record input primaries, white, encoding and domain. Establish the same for its output. A name such as “Cinematic 709” does not tell you whether the table expects LogC, S-Log, normalized Rec.709 or another representation. A correctly exposed shot in the wrong encoding still addresses the wrong samples.

A camera model is insufficient: curve generations and primaries may differ. Placement, a preceding shaper and any output rendering already applied also matter. Increasing table size cannot repair a wrong input. An undocumented table is an unknown transform, not a universal preset.

A simple table applies one function to each component. A 3×1D table provides a separate curve for R, G and B. Both remain channel-independent: output R depends only on input R. This suits independent tone responses, but cannot replace an arbitrary operation coupling the channels.

If output red must change with input green, an independent red curve is insufficient. A channel-mixing matrix or selective hue change generally requires another model. Increasing 1D length does not remove this structural limitation. A very long table of the wrong type remains the wrong type.

A uniform 3D LUT has N samples on each input axis. It contains N³ nodes, each storing an output triplet. A 17³ grid stores 4,913 triplets and a 33³ grid 35,937. This does not limit an image to seventeen levels: values between nodes are interpolated.

Node count and stored-value precision are different parameters. A large grid with coarsely rounded values can behave differently from a smaller adequately precise grid. Also compare format, axis order, domain and the application’s actual algorithm. Do not infer the row order of a text file from its appearance.

Linear 1D interpolation mixes two neighbouring outputs. Trilinear interpolation works along three axes using eight cell corners. Tetrahedral interpolation selects a tetrahedron within the cell and uses four vertices. Both need consistent handling of boundaries and exact node hits.

Do not choose a method on an “always better” promise. Error depends on the function, grid and test inputs. Methods can nearly agree on one transform and diverge on another. API labels such as BEST and DEFAULT are not fixed mathematical specifications either: state the method explicitly for reproducibility.

A uniform grid is uniform in a chosen coordinate, not necessarily in light or exposure stops. Over a broad linear range, shadows may receive too few samples. A shaper changes the coordinate used to address the table, redistributing the sampling density.

The table must be built for that coordinate. Arbitrarily placing log before a table constructed for linear input does not preserve the result. Baking samples the target through the inverse shaper; evaluation uses the corresponding forward shaper. Zero and negative values need a defined treatment rather than an arbitrary logarithm.

A finite table domain requires a policy outside it: clamping, extrapolation or another specified format/application behaviour. This lab explicitly clamps inputs to the boundary. The original value and an out-of-domain flag remain visible.

When several inputs collapse to one boundary output, a denser grid cannot recover them. An inverse table cannot determine which original value was lost either. Distinguish clipping at the LUT input, its output values and separate preview clipping. Values outside 0…1 are not automatically erroneous in a mathematical pipeline, but require a defined downstream policy.

A ramp reveals discontinuities and crude interpolation, but traverses only a small part of the RGB volume. A 3D transform needs coloured ramps, saturated areas, neutrals and independent points inside cells. Testing nodes alone is overly favourable: a correctly baked table matches the reference exactly there.

Compare the original processor and exported table with identical input, reader method and pipeline placement. Check floating-point outputs first, then final quantization and display. Avoid rebuilding from an already coarse table unnecessarily: a new filename does not remove accumulated error. Mean error does not replace inspecting the worst regions.

A technical table addresses a specific representation change. A creative table defines desired color and tone relationships. A display LUT may contain part or all of the viewing preparation. One table can combine these roles, so bypassing it does not necessarily remove only a “look”.

Do not reduce the intensity of an unknown technical transform as though it were an ordinary effect: mixing original and transformed images may mix numbers belonging to different representations. Separate technical stages from creative processing first. Establish whether the table includes output rendering and whether the color-management system repeats it. A wide working pipeline does not guarantee a wide domain in every inserted table.

An ordinary RGB → RGB table is pointwise: the same RGB input under the same settings produces the same output regardless of position or neighbouring pixels. It does not recognize faces, track motion or derive a content-dependent mask. Blur, denoising and local contrast require additional information.

A LUT can be applied through an external mask, but that does not store the mask inside the table. Likewise, CLF can carry a sequence of mathematical color operators without becoming a container for an arbitrary edit. Inspect the operation before baking and verify what the export actually contains.

Freeze the source processor and test inputs. Export the table, read it back using the recipient’s intended path and compare without creative compensation. Start with identity, black, white and neutrals, then examine coloured points and boundary neighbourhoods.

Separate transform error, file quantization and display output in the report. Matching one configuration does not validate every other mode. Before using ΔE, put both branches into a consistent colorimetric representation. A raw RGB difference is useful for implementation checks, but is not automatically a perceptual difference. Do not try to cure an incorrect input by selecting another interpolation method.

Deliver a short manifest with the table: input, output, domain, shaper, interpolation, version, purpose, limitations and expected placement. Include a known patch or frame before and after processing. A filename does not replace this record. The .cube extension does not guarantee identical interpretation in every program.

In Resolve, first identify the representation entering the intended node. Check the project’s technical and output transforms. Put the LUT in a clearly named separate node, compare against a reference and play the whole shot. Menus and supported formats vary; this chapter explains a checking procedure rather than promising identical commands everywhere. Do not repair a viewing fault by altering the table itself.