The Digital Canvas: Color Spaces, Bit Depth & RAW vs. JPEG

Photon capture and the photoelectric effect. A camera sensor is a grid of millions of microscopic light wells called photosites, each containing a silicon photodiode. When a photon of sufficient energy strikes the silicon, it liberates an electron — the photoelectric effect, the phenomenon for which Einstein won the 1921 Nobel Prize. During the exposure, each photosite accumulates freed electrons in proportion to how many photons landed on it. Modern back-side-illuminated sensors convert roughly 50–80% of incoming photons into electrons (their quantum efficiency). Each photosite has a full well capacity — typically tens of thousands of electrons on a full-frame sensor — and when the well is full, additional photons record nothing. That is physical, unrecoverable highlight clipping: the well simply cannot count any higher.

The sensor is colorblind. A photodiode counts photons; it does not know their wavelength. Color is manufactured by laying a mosaic of tiny red, green, and blue filters over the photosites — the Bayer color filter array, patterned 50% green, 25% red, 25% blue. The green bias is deliberate biomimicry: human photopic (daylight) vision, mediated by the cone cells of the retina, peaks in sensitivity near 555 nanometers — green — because the M and L cones that dominate our luminance perception respond most strongly there. Our rod cells, by contrast, handle dim scotopic vision, are essentially monochromatic, and peak nearer 498 nm, which is why moonlit scenes feel desaturated and blue-shifted (the Purkinje effect). The sensor's green-heavy sampling matches the channel our brains weight most heavily when judging brightness and detail. Because each photosite records only one of the three primaries, the missing two-thirds of every pixel's color must later be interpolated from its neighbors — a process called demosaicing.

Linear capture versus logarithmic perception — the central asymmetry of digital imaging. Silicon responds to light linearly: twice the photons, twice the electrons, twice the recorded value. Human vision responds roughly logarithmically (the Weber–Fechner relationship): we perceive equal ratios of luminance as equal perceptual steps. A jump from 1 to 2 candles feels like the jump from 100 to 200 candles — each is "one stop." This mismatch has a staggering numerical consequence. The analog voltage from each photosite is digitized by an analog-to-digital converter, typically at 14 bits in a modern RAW file, yielding 2¹⁴ = 16,384 discrete levels per channel. Because the encoding is linear and each photographic stop is a doubling, the levels distribute like this:

Fully half of all recorded tonal information lives in the single brightest stop, while the deep shadows — where the eye is exquisitely sensitive to gradation — are described by a handful of noisy values. This is the physical foundation of the "expose to the right" doctrine and of Chapter 2's rule that highlight recovery is cheap while shadow lifting is expensive.

Bit depth as gradation insurance. A JPEG stores 8 bits per channel: 2⁸ = 256 levels, for about 16.7 million nominal colors. A 14-bit RAW holds 16,384 levels per channel; when Lightroom promotes it into a 16-bit working precision, headroom rises to 65,536 levels per channel — over 281 trillion addressable colors. The extra levels are not visible on screen; they are editing headroom. Every strong curve, contrast move, or white-balance correction stretches some tonal ranges apart. Stretch 256 levels and gaps open between adjacent values; stretch 65,536 and the gaps stay far below the threshold of visibility.

What a RAW file actually contains. A .CR3, .NEF, .ARW, or .DNG is not an image. It is the mosaiced, linear, un-white-balanced sensor readout plus metadata: the black level, the white (saturation) point, the color filter response of that sensor model, the white-balance multipliers the camera suggested, and a small baked JPEG preview (which is what your camera's LCD and histogram actually show you — a critical deception covered in the recipe below). A JPEG, in contrast, has already been demosaiced, white-balanced, contrast-curved, gamma-encoded, converted to a small color space, quantized to 8 bits, chroma-subsampled (typically 4:2:0, meaning color resolution is stored at one-quarter of luminance resolution), and lossy-compressed via discrete cosine transform quantization. Every one of those steps discards information permanently.

Lightroom's engine: linear light inside enormous walls. When Lightroom's Develop module opens a RAW file, it demosaics the mosaic and places the result into its internal working space: a color space built on the ProPhoto RGB primaries with a linear (gamma 1.0) tone response, processed at high precision (16-bit and floating-point internally). ProPhoto RGB — originally Kodak's ROMM RGB — is deliberately gigantic. Its gamut covers on the order of 90% of the CIE 1931 chromaticity diagram of visible color, versus roughly 50% for Adobe RGB (1998) and roughly 35% for sRGB. In fact, ProPhoto's green and blue primaries lie outside the spectral locus entirely — they are "imaginary colors" no light can produce and no eye can see. That is not a flaw; it is engineering. By stretching the triangle beyond the visible horseshoe, every saturated color a camera sensor can register — a sodium-vapor streetlight, a backlit poppy, an LED-lit stage — fits inside the working space with room to move during editing.

A charming footnote: the histogram and RGB percentage readouts you see in the Develop module are not raw linear values. They are rendered in "Melissa RGB" — an unofficial Adobe space using ProPhoto's primaries but the sRGB tone curve — named after Adobe engineer Melissa Gaul. This is why Develop readouts appear as familiar 0–100% values rather than the punishingly dark linear data underneath.

Why editing in a small space destroys images: posterization and banding, quantified. Imagine a clear blue sky occupying a narrow tonal band — say 30 adjacent levels in an 8-bit file. Now apply a strong contrast curve that stretches that band across 100 output levels. The 30 input values cannot invent intermediate data; they land on 30 of the 100 output slots, leaving staircase gaps of 3–4 levels between neighbors. On a smooth gradient, the eye detects those steps instantly as concentric bands — posterization. The same move performed on 16-bit data inside ProPhoto has roughly 256 times the level density; the post-stretch gaps remain far below visual threshold. Similarly, if you convert to sRGB before editing, every color outside sRGB's small triangle is clipped to its boundary immediately: a saturated sunset's dozens of distinct orange chromas collapse into one flat wall of maximum orange, and no slider can ever separate them again. Lightroom's non-destructive, wide-gamut, high-bit-depth pipeline exists to postpone all such destruction to the final export — the only moment small spaces belong.

Non-destructive parametrics. Lightroom never alters your RAW. Every slider is a stored instruction; the pipeline re-renders from the original sensor data through the full instruction stack every time. This is why edit order in the interface is flexible even though the pipeline order is fixed (a fact that becomes strategically important in Chapter 5's Calibration section).

Recipe 1 — Capture settings for maximum canvas (do this once, today):

1. 01Set image quality to RAW (14-bit lossless compressed if your body offers a choice; avoid 12-bit and "lossy compressed" modes for critical work).
2. 02Set the camera's color space to Adobe RGB. This does not affect RAW data — it only tags the embedded JPEG preview — but it slightly widens the gamut of the histogram your camera shows you, making the in-camera clipping warnings marginally more honest.
3. 03Understand the deception: your camera's histogram and blinkies are computed from the JPEG preview, not the RAW. A "clipped" in-camera histogram often has a half stop to a full stop of genuine RAW headroom remaining. Learn your body's margin by deliberately overexposing a test scene in half-stop steps and checking recovery in Lightroom.
4. 04Expose to the right when dynamic range allows: bias exposure as bright as possible without clipping essential highlights, to spend your photons where the levels live (see the table in §1.1), then normalize brightness in Lightroom.

Recipe 2 — Export decisions by destination (the only time small spaces are correct):

Recipe 3 — Diagnosing banding you already have: if a sky shows bands after heavy editing, first check whether the source was JPEG (if so, the damage is baked; re-edit from RAW if it exists), reduce the aggression of Dehaze/Contrast in that zone, and as a last resort add fine luminance noise (Detail panel, Grain amount 5–15, size 25) — dithering that masks quantization steps with texture, the same trick audio engineers use.