Two thirds of the colour in your photo was calculated

One colour per photosite

A camera sensor measures light, not colour. Each photosite counts photons and reports a single number. To get colour, manufacturers lay a colour filter array over the sensor (the Bayer pattern, invented at Kodak in 1976) so each photosite sees only red, only green or only blue.

The pattern is half green, a quarter red, a quarter blue, in a repeating two-by-two tile: green and red on one row, blue and green on the next. Twice as many green sites, for the same reason green dominated the luminance formula two lessons ago: that is where your eye gets most of its detail.

So a 24-megapixel camera captures 24 million single-colour samples. It does not capture 24 million red-green-blue triples. Every pixel in your finished file has one measured channel and two that were calculated from the neighbours: a process called demosaicing. Two thirds of the colour data in every digital photograph is an interpolation.

This is not a scandal, and you would never notice it on a normal subject. You notice it on fine repeating detail: a distant brick wall, a tweed jacket, a window screen, the weave in a shirt. When the detail is near the sampling frequency of the array, demosaicing produces moiré (coloured wave patterns that were never in the scene) and occasionally false colour speckle along high-contrast edges.

The algorithms differ and the free tools let you choose, which most photographers never realise. RawTherapee offers AMaZE, DCB, LMMSE, IGV and others; darktable offers PPG, VNG4, RCD and LMMSE. AMaZE is usually the best on detail; LMMSE is noticeably better on very noisy high-ISO files; VNG4 suppresses maze artefacts in flat areas. Switching the algorithm on a difficult file is a free improvement available to nobody shooting JPEG.

What the camera throws away when it makes a JPEG

Shoot JPEG and the camera does all of the following, permanently, before you ever see the file:

1. Demosaics, with an algorithm you cannot change.
2. Applies white balance: bakes it in.
3. Applies a tone curve, usually a contrasty one.
4. Applies noise reduction and sharpening at settings you did not choose.
5. Converts from 12 or 14 bits per channel down to 8.
6. Compresses with JPEG’s lossy encoder.

A raw file is the sensor’s numbers before all six. It is typically 12 or 14 bits and, crucially, linear: proportional to the light, with no transfer function applied yet.

The two things raw gives you

White balance for free. In a raw file, white balance is still a set of multipliers waiting to be chosen. Shot under a tube light and got a green cast? Change one number in the developer and it is gone, with no cost at all. Fix the same cast in an 8-bit JPEG and you are stretching one channel against the others inside 256 levels, which is exactly the banding problem from two lessons ago.

Highlight headroom. Here the mechanism is worth knowing precisely. The three channels clip at different points, because white balance multipliers differ: under tungsten light the blue channel is multiplied up hard, under daylight the red is. When a bright sky is slightly overexposed, the green channel may be at maximum while red and blue still hold real values. A raw developer can reconstruct the clipped channel from the two that survived and recover the tone and colour. That is where the familiar “one to two stops of highlight recovery” comes from, and it is genuine.

There is also shadow room. 14 bits linear means 16,384 levels, distributed proportionally to light, so the shadows carry far more distinct values than an 8-bit gamma-encoded file does after it has been squeezed.

The costs, stated plainly

• Size. A 24 MP raw is 25 to 35 MB against 6 MB for a JPEG. A wedding at 2,000 frames is 60 GB rather than 12.
• Time. Every raw needs developing. For 40 product shots that is fine; for a school sports day it is an evening.
• Format churn. Raw formats are proprietary and differ per camera: CR3, NEF, ARW, RAF, plus phone DNG. New camera models are not supported by free developers until a release ships with the right decoding data, which can be months. Adobe’s free DNG Converter is the standard workaround: convert to DNG, and darktable and RawTherapee will open it.
• Nothing is “finished”. A raw file always looks flatter than the camera’s JPEG on first open, because the camera’s contrast and saturation were not applied. Beginners conclude raw is worse. It is not finished, which is the point.

The free path is strong here. darktable and RawTherapee are both mature, both free, both run on a modest laptop, and both do things Lightroom does not: RawTherapee’s demosaic choice and darktable’s profiled denoise are best-in-class. ART is a friendlier fork of RawTherapee for people who find it dense. On a phone: most mid-range Android models shoot DNG in Pro mode and iPhone Pro models shoot ProRAW, and Lightroom mobile’s free tier develops both.

What raw does not do

Raw fixes exposure, white balance and a stop or so of highlights. It does not fix focus, it does not fix motion blur, and it does not fix a highlight where all three channels are at maximum: there is nothing left to reconstruct from, and what the developer paints in there is a flat guess. A blown sky is a blown sky in any format.

It also does not rescue a composition. The commonest mistake among people who have just switched is shooting carelessly on the assumption that raw covers it. It buys you latitude, not a different photograph.