The image pipeline

An image pipeline needs four separate descriptions: what the values represent, their primaries and white point, their encoding, and how they will be rendered.

What the values mean

Scene-referred values describe captured or simulated scene light. They may be linear or log-encoded. Display-referred values describe an image prepared for an output condition; they can also be stored in linear form.

“Linear” and “scene-referred” are therefore not synonyms. Neither JPEG nor ProRes alone specifies the complete color interpretation of an image.

  1. Scene lightCapture or rendering
  2. Working valuesDecode → process → render for output
  3. Display lightOutput encoding → display response
Figure 3.1. Reference and encoding. Schematic. Either reference domain can use linear or nonlinear encoding. Source.

An exposure adjustment multiplies scene-linear values by 2 to the power of the stop change. Apply it before the view transform when that is the intended operation. A display brightness adjustment is a different operation.

RGB requires primaries, a white point, and an encoding. HSL and HSV reorganize encoded RGB values into convenient controls; their lightness/value coordinates are not perceptually uniform.

Y′CbCr separates a weighted sum of nonlinear R′G′B′ (luma, Y′) from chroma differences. Luma is not linear luminance Y. Matrix coefficients, signal range, and chroma sampling are part of the interpretation. Analog YUV and digital Y′CbCr are related concepts, not interchangeable names.

  • RGB / HSL / HSVRGB components or rearrangements of them
  • Y′CbCrEncoded luma and chroma differences
  • CIELABApproximate lightness and opponent coordinates
Figure 3.2. Coordinates have different jobs. Schematic. Coordinates need a defined space and viewing assumptions. Source.

Choose a representation by the operation you need:

Representation Useful for What it does not establish
Linear RGB with named primaries and white Adding light, compositing, primary conversion A perceptually uniform editing space
Encoded RGB such as sRGB or LogC4/AWG4 Storage, exchange, or specified grading operations Scene/display meaning from “RGB” alone
XYZ / xyY Standard-observer measurement and chromaticity Spectrum or complete appearance
Lab / Luv Approximate perceptual comparison under specified reference conditions An observer-independent visibility threshold
Y′CbCr Video transport and chroma subsampling Linear luminance or a camera RAW representation
ICtCp The HDR signal representation specified in BT.2100 A replacement name for every Y′CbCr signal
HSL / HSV Convenient controls derived from RGB Equal lightness or brightness at equal L/V
CMYK with a printing condition/profile Ink separation and print reproduction One universal four-channel color space

The definitions of television signal representations come from BT.709 and BT.2100; device interpretation is the role of ICC profiles. For example, CMYK percentages depend on the inks, substrate, and separation. They are not a fixed inverse of RGB.

In 4:2:2, chroma has half the horizontal sampling density of luma; in 4:2:0, it has half in both directions for a progressive image. Chroma siting and filtering still matter. This sampling reduction can affect edges and keys even when a flat patch looks correct. The appendix gives the BT.709 matrix and range conventions.

Encoding and rendering

An OETF maps scene light to an encoded signal; an EOTF maps a signal to display light. The overall scene-to-display relationship is the OOTF. They need not be simple inverses once rendering intent and viewing conditions are included.

sRGB uses a linear segment near black and a power segment above it. A simple gamma power law is only an approximation. This explorer compares the equations; it does not simulate monitor brightness.

Linear, sRGB, and a power lawEncoded signalLinear input000.250.250.50.50.750.7511
  • Linear
  • sRGB
  • Gamma 2.2

At 0.18: sRGB 0.461; gamma 2.2 0.459.

Figure 3.3. Linear, sRGB, and a power law. Encoding curves over [0,1]; no display-luminance prediction. Source.

PQ is an absolute display-luminance encoding, defined up to 10,000 cd/m². HLG starts with a relative scene-light encoding and includes a system relationship to the display. Their independent variables differ; overlaying them without defining normalization is misleading. ITU-R BT.2100-3.

Tone mapping compresses an intensity range. Gamut mapping handles colors the target cannot reproduce. A useful view transform coordinates both; independently clipping RGB channels can change hue and erase distinctions.

Compression keeps highlight distinctionsMapped outputScene-linear input0020.2540.560.7581
  • Clip
  • x / (1 + x)

Exposure multiplier 1.00×. At input 4, compressed output 0.800.

Figure 3.4. Compression keeps highlight distinctions. Illustrative x/(1+x) versus clipping. Neither curve represents AgX, Filmic, or ACES. Source.

Blender’s AgX and Khronos PBR Neutral serve different rendering priorities. AgX handles a broad exposure range and desaturates bright highlights; PBR Neutral targets faithful base-color reproduction for well-exposed products under neutral illumination. Neither should be described as literal film chemistry. Use the actual versioned implementation for image comparisons, rather than labeling an arbitrary S-curve with a product name. Blender, Khronos.

A transform changes values; a tag describes them

Assigning the correct source interpretation tells a pipeline what the existing numbers mean. Converting changes the numbers to preserve a specified relationship in a new representation. Assigning sRGB to LogC data does not perform a LogC-to-sRGB conversion.

For a typical camera log input, the conceptual sequence is decode the camera transfer → convert linear coordinates, including the specified white adaptation → work on the image → render for the target → encode the output. Some applications combine these steps. Record their domains so that a camera LUT, managed timeline, and viewer do not each apply the same rendering.

Chromatic adaptation is especially important when converting D65 camera/display coordinates to the ACES white. It models a change in adopted white; it is not simply dividing the XYZ components by a new white or measuring the scene’s illuminant. Use the adaptation specified by the input transform, not an extra correction added by eye. ARRI LogC4 conversion to ACES, §4.3.2

A useful first diagnostic is an exposure ramp in scene-linear values. Move it by one stop before rendering, then repeat the adjustment on the encoded output. The operations generally differ because rendering and encoding are nonlinear. This explains why operation order belongs in the workflow specification.