What the sensor kept. What processing baked in.

A photosite converts absorbed radiation into an electrical response. Illumination spectrum, object properties, optics and channel spectral sensitivities all contribute. Two stimuli that match for a human observer need not produce identical channel ratios in a particular camera. Neutral and coloured targets therefore test different aspects of reproduction.

Measurements have limits: sensor saturation, noise and later numerical precision. Once a channel reaches the physical limit, a wider gamut or a different log curve cannot restore its original value. The lab deliberately places a normalised saturation limit of one before both branches. It demonstrates irreversibility, not a particular camera specification.

In a common single-chip colour-filter-array sensor, each photosite measures one spectrally weighted component, not a complete RGB pixel. Bayer is common, but it is not the only filter arrangement. Do not apply its layout automatically to every camera, RAW type or stored format.

Producing a full-colour image requires reconstructing missing spatial components. This differs from Y′CbCr subsampling of already developed video: both involve spatial samples, but at different stages and with different meanings. A RAW mosaic is not simply an undeveloped 4:2:0 video file.

RAW is a family of ways to retain data close to capture and defer some interpretation. It does not guarantee that no processing has occurred. A format may include corrections, compression, nonlinear encoding or part of reconstruction. Read the specific format and decoder documentation.

Blackmagic RAW, for example, divides processing between camera and software, including a partial demosaic. Metadata provides adjustable decode settings, but those settings cannot return information absent from the recording. Sensor saturation, severe noise and losses from a particular compression method do not disappear because a file is called RAW.

Demosaicing estimates missing components using neighbouring measurements. The algorithm affects fine lines, coloured boundaries and false detail. It cannot know the scene perfectly: repetitive patterns can yield ambiguous observations. Later sharpening may emphasise a reconstruction error rather than useful texture.

After decoding, the image has definite values and a colour representation. Access to RAW controls is not the same as the state reaching the node graph. For a controlled comparison, fix decoder, version, quality and output settings. Different applications can display the same RAW differently without modifying the source file.

In a simplified linear model, white balance changes relative channel gains. A real RAW decoder combines this with camera calibration and illuminant interpretation; Kelvin and Tint are not universal RGB coefficients across camera models. Changing a balance value does not change the spectrum that originally illuminated the object.

When original samples remain available, a decoder can apply a different decision without working backwards from an already limited result. Developed video may have combined the old balance with nonlinear processing, clipping and quantisation. Inverse gain cannot restore a clipped channel. RAW cannot undo saturation that already happened in the sensor either.

Sensor channels should not automatically be treated as Rec.709, AWG or S-Gamut primaries. Calibration and processing translate camera response into a defined representation. Even when a stage is matrix-based, the entire camera colour pipeline is not one universal matrix for every illuminant.

Outputs may include a technical representation, a viewing rendering and a creative style. Distinguish them. A technical input transform need not immediately produce the desired contrast, while a marketing look does not replace the source definition. Establish the correct data state before choosing the artistic result.

The transfer function specifies encoding of light values; primaries and white point define the channels colorimetrically. A curve may accompany different gamuts. S-Log3, for example, can accompany S-Gamut3 or S-Gamut3.Cine: selecting only the correct log name is insufficient.

Check the generation as well. LogC3/AWG3 and LogC4/AWG4 are not interchangeable labels. Also determine whether the decoder has already transformed the image before your CST. The input profile must describe incoming values, not the original camera mode before intermediate processing.

RAW describes a format and deferred interpretation of measurements; log describes value encoding. Nonlinear encoding can exist inside a RAW format, and a RAW decoder can output log-encoded imagery for further work. “RAW or log?” is therefore incomplete without identifying whether you are comparing recording modes or decoder outputs.

A developed log file may already be demosaiced, balanced and stored as RGB or Y′CbCr with chosen sampling and compression. It remains gradeable, but the reversible operations differ. Do not assume that a log intermediate retains the original RAW camera controls.

For ARRI, distinguish ARRIRAW from developed LogC recording, then identify LogC3/AWG3 or LogC4/AWG4 and decoder interpretation. For Sony, check not only S-Log3 but the accompanying gamut, recording mode and whether a viewing look was baked in. For Blackmagic RAW, identify colour-science generation, decode settings and any sidecar.

This is a checklist, not a universal camera-compatibility table. Features, menus and recorded metadata depend on model and version. When handing off footage, include exact recording and decoder-output descriptions rather than “everything was shot in log”.

Start with Camera RAW where the format exposes it, and with project colour management. Establish which operations belong to the decoder, RCM or ACES, and which remain in nodes. Supported fields depend on format; a missing control does not prove that the image is technically incorrect.

Do not add a habitual input CST on top of a transform that already occurred. Use a separate version, a known source and neutral patches to diagnose. Document three boundaries: decoder output, grading input and final output. This exposes double normalisation through the processing path rather than hiding it with contrast.

First bring each verified source pair into a consistent working process. Then compare exposure, neutrals, skin and saturated objects under the same viewing conditions. A meaningful comparison is impossible when one camera has already received display rendering and another has not.

Correct input handling does not guarantee identical images: spectral sensitivities, lenses, lighting and processing can leave differences. Neutral balance does not fix every coloured object simultaneously. Apply restrained targeted adjustments and inspect several frames instead of building a universal match from one patch.

Inverting a gain is not the same as recovering lost information. A known transform that did not clip can be inverted within numerical precision. When many original values collapse to one, there is no unique inverse. Highlight reconstruction can use channel correlations and estimates without proving exact recovery of the original spectrum.

Distinguish sensor saturation from limits introduced by decoding or display. Change the relevant stage temporarily in a test version and check whether genuine detail differences remain. Do not promise recovery of every overexposure merely because the footage is RAW or log.