Designing colour for a screen you will never see
A print designer can hold a proof. A product designer can hold a moulded part. An interface designer sends a specification into a population of devices they will never touch: a five-year-old laptop with a dim panel, a phone with an aggressive saturation mode, a television in a shop with every enhancement switched on, a desktop monitor whose owner turned the brightness down because it hurt their eyes. The same hexadecimal value produces genuinely different colours across that spread, and no amount of care at the design stage changes it.
Interfaces and games therefore treat colour as something to be made robust rather than exact. The goal is not that everyone sees an identical colour, which is unachievable, but that the meaning carried by colour survives the variation.
What a colour value on a screen actually means
When a stylesheet specifies a colour, that value is interpreted according to a colour space, and the near-universal default on the web is sRGB, a standard defined in the 1990s around the behaviour of cathode ray tube monitors. Its transfer function is roughly a power law with an exponent near 2.2, which means the numeric values are not proportional to emitted light: doubling the number does not double the photons.
Two consequences follow, and both surprise people regularly.
- Blending in the wrong space darkens things. Averaging two sRGB values numerically is not the same as averaging the light they represent. Gradients between saturated complementary colours often develop a muddy dark band in the middle for exactly this reason. Blending in a linear space, where values are proportional to light, avoids it.
- Alpha compositing inherits the same problem. Semi-transparent overlays, antialiased text edges and soft shadows all involve blending, so a system that composites in a non-linear space produces edges and fades that are subtly too dark.
- Wide gamut is now real. Many current displays cover the P3 gamut rather than sRGB, and modern platforms let content request it explicitly. A colour that was the most saturated available red in sRGB is no longer the reddest thing on the screen.
- Colour management is partial. Browsers and operating systems handle tagged images reasonably well and untagged content by assumption. An untagged photograph on a P3 display can render noticeably more saturated than intended, which is why exporting assets with an explicit profile still matters.
Choosing colours that behave
Most design systems were built on HSL, which arranges colour by hue, saturation and lightness. It is easy to reason about and it lies: two colours with the same HSL lightness can differ enormously in perceived brightness, because the model takes no account of the fact that the eye is far more sensitive to green than to blue. A palette generated by rotating hue at fixed lightness produces a yellow that glares and a blue that vanishes.
Perceptual spaces and why they help
Perceptual colour spaces fix this. CIELAB and its cylindrical form LCH, and more recently the OKLab family and its OKLCH form, are constructed so that equal numeric steps correspond to roughly equal perceived steps. Setting a palette in OKLCH gives hues that genuinely match in lightness, and generating a tint ramp by stepping lightness produces even progressions instead of the bunching that plagues HSL ramps.
- Anchor the ramp on lightness. Define each step of a scale by its perceptual lightness, then choose hue and chroma within it. This keeps the whole system consistent when a new brand colour is added later.
- Cap chroma per lightness. Very light and very dark colours cannot be highly saturated in any real gamut. Generating without a clamp produces values that get silently clipped, collapsing several intended steps into the same rendered colour.
- Keep semantic tokens separate from raw values. A token named for its role — surface, border, danger — can be retargeted for a new theme. A token named for its colour cannot, and every system that shipped a variable called blue-500 eventually had to make it green.
- Test at the extremes. Check the palette on a dim panel, on a phone in sunlight, and with the operating system contrast setting raised. Colours that only work at the designer’s brightness are not finished.
Dark mode is not an inversion
Producing a dark theme by inverting a light one reliably fails. Saturated colours that read as confident on white become vibrating and harsh on near-black, because the perceived contrast between a saturated hue and a dark ground is much higher. Shadows, which convey elevation on a light surface, become invisible and must be replaced by lighter surfaces or subtle borders.
Practical dark themes usually raise the lightness and lower the chroma of accent colours, avoid pure black in favour of a very dark neutral with a slight hue, and rebuild the elevation system rather than translating it. They also account for the fact that dark mode is often used in dark rooms, where the eye is partly dark-adapted and a bright element is genuinely uncomfortable rather than merely prominent.
Games: a rendering pipeline with a colour problem at each end
Real-time rendering has the same linear-versus-encoded issue as interface work, but with far more at stake because the whole point is simulating light transport.
The linear workflow
The standard modern pipeline is a linear workflow. Texture assets that represent colour are authored in sRGB and converted to linear when sampled. All lighting, shading and blending happens in linear space, where adding two light contributions is physically meaningful. The result, which can span a very wide range of values, is then compressed by a tone mapping operator into a displayable range and encoded back for the target display.
- Not every texture is colour. Normal maps, roughness maps and metallic masks store data, not colour, and must never be passed through an sRGB conversion. Doing so is one of the classic causes of subtly wrong shading that nobody can locate.
- Tone mapping defines the look. The choice of operator — a filmic curve, a variant of the ACES transform, a custom artist-driven curve — determines how highlights roll off and how saturated bright areas behave. It is the game equivalent of a film’s output transform, and studios treat it as an art direction decision rather than a technical default.
- Exposure and adaptation are simulated. Automatic exposure that adapts as the player moves from a dark interior to bright daylight mimics the eye, but if it is too aggressive it makes the world feel unstable, and if it is too slow it strands the player in a white or black frame.
- Bit depth and banding. Smooth gradients in skies and fog reveal quantisation steps in an eight-bit output. Adding a small dither before quantisation is cheap and effective, and it is the same trick that film delivery uses.
HDR output and the calibration screen
Games were among the first consumer software to ship high dynamic range output widely, and they inherited all its difficulties. A game rendering internally in linear floating point can map to a standard display or to an HDR one with a peak brightness that varies by several multiples between panels, and the platform provides only partial information about what the connected screen can do.
Hence the calibration screen almost every HDR game shows at first launch: adjust until the logo is barely visible, set the peak brightness until detail in the bright element just disappears, set the paper white level for menus and subtitles. These screens are asking the player to supply the display characterisation the system cannot reliably query. They are frequently poorly explained, and a substantial share of players end up with a badly configured picture and conclude that HDR looks bad.
- Show a real scene, not an abstract pattern. Calibration steps that use gameplay imagery are set correctly far more often than ones showing a grey ramp with an instruction.
- Separate paper white from peak. Interface elements and subtitles at scene peak brightness are painful. They need their own, much lower, level.
- Let it be redone. Players change screens and rooms. A calibration buried in a first-launch sequence with no way back is a permanent misconfiguration.
Colour as information in play
Games use colour to carry rules. Item rarity is signalled by a colour scale, team affiliation by hue, damage type by a palette, interactive objects by an accent that appears nowhere else in the environment art. This is efficient and instantly learned, and it is also fragile.
The fragility has three sources. Colour vision varies across the population, so any distinction carried by hue alone excludes some players; this is well covered by accessibility guidance and by the colour-blind modes now common in large titles, and the durable fix is redundancy — shape, icon, position or text alongside the hue. Environment lighting varies, so an accent colour tuned in a neutral test level can be swallowed by a heavily tinted scene. And displays vary, so a subtle distinction between two similar hues survives on a reference monitor and collapses on a cheap panel.
The robust approach is to make gameplay-critical colour distinctions large in perceptual terms — well separated in lightness as well as hue — to reserve a small set of colours exclusively for interactive meaning, and to keep environment art from competing with them. Studios that treat the gameplay palette as a protected resource, rather than as something art direction can borrow from, ship games that stay readable in their own most spectacular scenes.
The through-line across interfaces and games is the same. Screen colour cannot be controlled, only anticipated. Systems built on perceptual models, with generous margins, redundant signals and honest testing on bad hardware, degrade gracefully. Systems built on exact values chosen on one excellent monitor do not.
