AI Glossary / Computer vision and multimodal AI / Beginner
Tone curve
Definition
A tone curve is a mapping from input image values to output values, used to adjust brightness and contrast across different tonal ranges. It is commonly displayed as a graph with input values along the horizontal axis and output values vertically. Changing the curve changes how shadows, midtones, and highlights are represented.
What it means
A straight diagonal line represents an identity mapping: each input value keeps the same output value. Raising part of the curve brightens the affected values, while lowering it darkens them under the usual RGB brightness convention.
The slope matters as much as the curve's height. A steeper section spreads nearby input values farther apart, increasing their output separation. A flatter section brings them closer together. A completely flat section maps several inputs to the same output, losing their distinction in the rendered result.
The numbers must be interpreted in their working color space and encoding. A curve applied to linear-light values is not equivalent to the same-shaped curve applied to gamma-encoded values. A tone curve is a numerical transformation, not a universal measure of physical brightness.
How it works
An editor can define a curve through control points, a mathematical function, or a lookup table. Interpolation supplies mappings between specified points. A global curve applies the same input-output rule throughout the affected image or selected region.
Curves may operate on individual color channels or a brightness-related component. Adjusting channels differently changes color balance. Applying nonlinear functions independently to RGB channels can also affect color relationships, so the chosen processing mode matters.
A monotonic curve preserves tonal ordering: larger inputs do not become smaller outputs. A nonmonotonic curve can reverse ordering in part of the range and create deliberate special effects. Monotonicity alone does not preserve every detail, because clipping, flat portions, and finite output precision can still merge distinct input values.
Three practical examples
1. A product photographer uses a curve to lift midtone detail while keeping bright reflections controlled. The image is checked against the intended product appearance rather than accepted solely because it looks more dramatic.
2. A designer applies a consistent tonal style to a group of images. Each result is reviewed because differently exposed source images can respond differently to the same mapping.
3. An imaging pipeline uses a lookup-table curve for repeatable rendering. Test patches verify that expected input values produce the intended outputs and that important tonal distinctions are retained.
Think of it this way
Think of a translator that assigns every number on one scale a number on another. Changing the translation stretches some intervals and compresses others. The analogy explains value mapping. It is limited because perceived image contrast also depends on neighboring content, viewing conditions, color encoding, and display behavior.
Worked example
Define a simple piecewise-linear curve with normalized control points (0, 0), (0.25, 0.15), (0.75, 0.85), and (1, 1). Join neighboring points with straight lines for this teaching example.
Between input 0.25 and 0.75, the slope is (0.85−0.15)/(0.75−0.25) = 1.4. Inputs 0.40 and 0.60 therefore map to 0.36 and 0.64. Their separation grows from 0.20 to 0.28.
In the shadow interval from zero to 0.25, the slope is 0.15/0.25 = 0.6. Inputs 0.05 and 0.15 map to 0.03 and 0.09, reducing their separation from 0.10 to 0.06.
This curve increases midtone separation while compressing the ends of the range. It does not increase contrast equally everywhere. Real editing tools often interpolate smoothly rather than using these straight segments, so their exact intermediate values may differ even with similar control points.
Illustrative case study
Fictional case study: CatalogLight prepares photographs of dark woven bags. A contrast curve makes the overall collection look punchier, but some fine weave differences disappear in the darkest areas.
The editor reviews the curve's shadow behavior and reduces that compression. A separate check examines highlights on metal fittings. The team compares final images at their actual website size, where subtle changes can be less visible than in an enlarged editing view.
The approved curve is saved with the processing record, but each source remains available for individual adjustment. This fictional example illustrates tonal tradeoffs; it does not imply that one curve can make all images equally accurate or that darkening always improves presentation.
Common misconceptions and mistakes
Do not assume an S-shaped curve increases contrast in every tonal range. Its slope can increase separation in one range while compressing another.
Do not confuse moving image values with recovering clipped capture information. If the original measurement already lost distinctions, remapping the remaining value cannot recreate them reliably.
Do not treat the same numeric curve as interchangeable across color spaces or processing stages. Input encoding and previous adjustments affect what its values represent and how its output will look.
Business use
Tone curves support consistent image rendering, controlled contrast, and creative style. They are useful because a compact mapping can be saved, inspected, and applied repeatedly.
Evaluate the ranges that matter to the product: dark fabric, pale labels, reflective details, or subtle gradients. Check for clipping, banding, and unexpected color shifts after export. Keep the original source and an editable adjustment when possible so a later presentation change does not require reconstructing values from an already compressed or clipped result.
People also ask
Does raising a curve always affect the entire image equally?
No. It affects values according to the changed portion of the mapping, and any mask or selected processing mode also limits the effect.
Can a curve change colors?
Yes. Channel-specific curves directly change color balance, and nonlinear RGB processing can alter color relationships.
Is an identity curve always a diagonal line?
With equally scaled input and output axes using the same convention, yes: output equals input. Display conventions should still be checked.
Check your understanding
A curve has a steeper middle section but flatter shadows. Does it preserve or increase all original tonal separations?
Answer: No. It expands separation in the middle and compresses it in the shadows. The effect depends on where the input values lie.
Related terms
Learn next
Plot input-output pairs for a simple curve, then examine its slope in shadows, midtones, and highlights. Learn the difference between linear-light and encoded values before comparing curves from different image-processing stages.
More in Computer vision and multimodal AI
- Texture - Texture is the visual pattern formed by variations in an image region, such as the grain of wood, a fabric weave, or a speckled surface.
- Texture synthesis - Texture synthesis creates new image content with the visual character of a texture, such as woven fabric, stone, or scattered leaves.
- The Bitter Lesson - The Bitter Lesson is Richard Sutton's 2019 argument that broadly applicable methods able to use increasing computation, especially search and learning, have...
- Train/validation/test split - A train/validation/test split divides a dataset into parts with different roles: training fits a model, validation guides development choices, and testing...
- Transcript-based editing - Transcript-based editing is a way to edit recorded speech and its associated media by selecting and arranging words in a time-linked transcript.
- Trifocal tensor - The trifocal tensor is a mathematical representation of the projective geometry connecting three camera views.
Sources
Adobe, Adjust Color and Tone with Curves: https://helpx.adobe.com/photoshop/desktop/adjust-color/color-corrections/adjust-color-and-tone-with-curves.html
The official reference supports curve axes, tonal adjustment, and channel controls. The piecewise-linear calculation and fictional catalog workflow are original teaching examples.
Previous: Tolerance Next: Tone mapping