Is a LUT an AI model?

A conventional LUT is a stored mapping. AI may generate or select one, but lookup itself does not imply machine learning.

What does “3D” mean in a color LUT?

It refers to three input coordinates, usually red, green, and blue.

Can one LUT work correctly on every image?

Only if the images and workflow meet its assumptions; identical pixel numbers can mean different things in different color spaces.

AI Glossary / Computer vision and multimodal AI / Beginner

Look-up table

Definition

A look-up table, commonly called a LUT, stores output values that can be retrieved using an input value or combination of values. In image processing, it often maps pixel intensities or colors to new values, replacing repeated calculations with table access and, where needed, interpolation.

What it means

Suppose an operation always converts input intensity eighty to output intensity one hundred. A table can store that answer once and reuse it wherever eighty appears. The table describes a mapping; it does not need to understand what the image contains.

A one-dimensional LUT takes one input coordinate. Separate one-dimensional curves can adjust red, green, and blue independently. A three-dimensional color LUT uses the combination of red, green, and blue as its coordinates, allowing an output channel to depend on all three inputs.

The name also has broader computing uses. Here the emphasis is on images, where LUTs support tonal adjustments, color transformations, visualization palettes, and repeatable looks.

How it works

Define the mapping's input domain and the output values at sampled positions. For a complete discrete table, the input can directly select an entry. For a sparsely sampled continuous mapping, locate nearby samples and interpolate according to an agreed method.

In a simple unsigned eight-bit intensity mapping, a table can contain one entry for each of the 256 possible input values. OpenCV's tutorial demonstrates this style of image transformation.

Color workflows can use three-dimensional grids and supporting shaper transforms that redistribute input values before lookup. The input color space, output color space, interpolation, and treatment of out-of-range values all affect the result. A LUT file without that context is an incomplete description of the intended processing.

Three practical examples

1. Preview rendering: A video team uses a documented LUT to preview a particular input encoding on a target display. It checks that the source footage matches the LUT's expected input.

2. Scientific visualization: A grayscale measurement image is assigned a color palette through a LUT. The displayed colors help distinguish ranges, while the original numerical measurements remain available for analysis.

3. Repeated tonal adjustment: An image service precomputes a fixed curve for eight-bit thumbnails. Every pixel with the same input receives the same mapped value, making the operation reproducible across images.

Think of it this way

Think of a conversion chart with a column for the value you have and another for the value you need. Looking up an answer can be simpler than recalculating it every time.

The analogy becomes incomplete for multidimensional color tables: the answer depends on a combination of coordinates, and values between stored points require interpolation. The chart also cannot decide whether its conversion is appropriate for a particular photograph.

Worked example

Consider a one-dimensional table with these sample pairs: input 0 gives output 0, input 64 gives output 80, input 128 gives output 150, and input 255 gives output 255.

For input 96, assume linear interpolation. The input lies halfway between 64 and 128 because (96 − 64)/(128 − 64) = 0.5. The output is therefore 80 + 0.5 × (150 − 80) = 115.

This is an interpolation example, not a complete eight-bit table. A fully populated version could store the answer for every integer input directly. Different rounding or interpolation rules could change results near sample boundaries.

Now consider storage for a three-dimensional LUT with seventeen samples along each input channel. It contains 17³ = 4,913 grid points. Storing three output channels per point requires 14,739 scalar values, before file overhead. The cube refers to color coordinates, not three-dimensional scene geometry.

Illustrative case study

Fictional case: An online furniture retailer distributes a preferred color LUT to several photography contractors. One contractor applies it to already-rendered images, while another applies it to a different camera encoding. The resulting product colors do not agree.

The retailer documents the expected input space and supplies a reference image with an approved output. Each contractor tests the complete conversion before processing the catalog. The team also checks neutral and saturated samples rather than judging only an attractive hero image.

The scenario illustrates the importance of consistent interpretation. A shared LUT alone does not guarantee color accuracy or compensate for inconsistent capture conditions.

Common misconceptions and mistakes

Do not assume every LUT is a creative filter. Some implement technical conversions, others implement artistic changes, and some combine both.

Do not assume a three-dimensional LUT can perform spatial edits such as selective sharpening or removing a person. A fixed color mapping has no access to object identity or neighboring pixels unless it is embedded in a larger, context-dependent system.

Avoid claiming reversibility without checking the mapping. Clipping or mapping several inputs to one output loses distinctions. More grid samples can improve approximation, but they do not correct an inappropriate transform or incorrect input assumptions.

Business use

LUTs help organizations distribute consistent image transformations across compatible tools. They can also reduce repeated computational work, although actual speed depends on hardware, data access, table size, and implementation.

Manage them as versioned production assets. Record their purpose, expected input and output, interpolation requirements, and known limitations. Test gradients and representative colors for artifacts, and verify that export software applies the transform exactly once. For quantitative imagery, distinguish the displayed palette from the underlying data so presentation choices do not become mistaken measurements.

Check your understanding

A LUT maps both input 250 and input 255 to output 255. Can its output uniquely reveal which input occurred?

Answer: No. That mapping has discarded the distinction between those inputs.

Related terms

Color space
Transfer function
Log encoding
interpolation
color management
quantization
colormap

Learn next

Build a small one-dimensional table and compare direct lookup with interpolation. Then examine a color LUT's input specification and test a neutral ramp before using it in a complete image workflow.

Sources

OpenCV, How to Scan Images, Lookup Tables and Time Measurement: https://docs.opencv.org/4.11.0/db/da5/tutorial_how_to_scan_images.html

OpenColorIO, LUT Baking Tutorial: https://github.com/AcademySoftwareFoundation/OpenColorIO/blob/main/docs/tutorials/baking_luts.rst

Both official references were opened. The mapping, storage calculation, and retailer scenario are original illustrations.

Previous: Look-ahead (backtracking) Next: Lookahead in regex