Beyond RGB. Building an RGBAL architectural light with WLED.

This project examines whether WLED, a platform generally associated with decorative LED strips, can form the control basis of a high-quality indoor architectural luminaire. The central challenge is not producing five PWM outputs. It is determining, accurately and repeatably, how those outputs should be combined.

WLED provides networking, presets, effects, a well-developed user interface and a practical route to an ESP32-based prototype. Current WLED supports analogue RGB, RGBW and five-channel RGBCCT outputs, as well as addressable RGBW and selected addressable RGBCCT chipsets.¹ This makes the electrical control of a five-output experiment comparatively straightforward.

The limitation is the colour model. Stock WLED stores an RGB colour, one white value and a separate CCT blend. It does not have a native five-channel colour model for red, green, blue, amber and lime. A basic RGBAL prototype can operate from stock WLED, but the firmware does not know how to mix five measured spectra into a particular white, pastel or saturated colour.

The project described here extends that capability. The sdgelectronics/WLED_RGBAL fork adds four- and five-primary PWM buses, a spectrometer-derived lookup table, 16-bit internal channel data and a calibration API. The software and simulation work are complete. Closed-loop measurements on the finished hardware remain outstanding.¹³

Every extra letter normally means another independently driven emitter colour. The word “normally” matters: LED naming is not standardised consistently across strips, fixtures and marketplaces. A part number and a channel diagram are more trustworthy than the acronym printed in a listing.

So why combine amber and lime?

RGBAL gives the mixer more choices. An amber target can use the amber emitter rather than wasting output in a red-plus-green compromise. A near-white or pale colour can lean on lime for efficient mid-spectrum energy, while amber helps warm objects and skin. The extra degrees of freedom can be used to optimise brightness, colour rendition, electrical load or thermal headroom.

What it does not do automatically is make the outer colour gamut larger. Amber and lime may sit inside the triangle formed by the RGB primaries on a chromaticity diagram. Their value is often spectral quality and efficiency inside that gamut, not simply a larger polygon.

That last point is the reason multi-primary lighting exists. Two light sources can land on nearly the same chromaticity and CCT while producing very different spectral power distributions. Objects then reflect those spectra differently, so one source can make timber, red paint or skin look lively while the other makes it look flat.

CRI is still a familiar shorthand, but it compresses a complicated subject into one average. TM-30 adds fidelity (R_f), gamut (R_g) and hue-specific information. The US Department of Energy’s review found that TM-30-based criteria captured preference better than the older single-number approaches in the studies it analysed.¹⁰

WLED’s documented light capabilities are RGB, white and CCT. Its JSON colour array contains three or four bytes for RGB or RGBW, while CCT is held as a separate value for each segment.³ That works well for RGBCCT. The fourth value sets the amount of white, and the CCT value divides it between the warm-white and cool-white emitters.

The official compatibility page dates analogue single-colour, CCT, RGB, RGBW and RGBCCT support to WLED 0.13.1. It also lists addressable WS2805 and FW1906 RGBCCT support from 0.15.0-b2.¹ Those are genuine five-channel lighting modes, but they still use the RGB-plus-white-and-temperature model rather than a general five-primary colour vector.

The current PWM bus code makes this explicit. The five-channel type is labelled PWM RGBCCT. It writes RGB directly, calculates the warm-white and cold-white values, then sends them to channels four and five.⁴

What happens if WW and CW become amber and lime?

Electrically, it works. Map the nominal warm-white output to amber and the nominal cold-white output to lime. The WLED white slider now controls how much amber/lime contribution is requested, while the CCT slider moves the balance from amber to lime.

For a prototype, this is useful, but it is still an approximation. WLED continues to treat the two channels as a white-temperature pair. Auto-white can add them whenever it extracts white from RGB. Effects also tend to generate RGB values, so they may not use the extra pair in the way a five-primary mixer would. The CCT additive-blending setting changes how strongly both channels can be driven near the centre. WLED notes that high additive blending can increase heat and reduce LED life.²

At a fixed temperature and current, additive light is conveniently linear. Measure the CIE XYZ contribution of each emitter and place those measurements into a matrix. A requested target can then be expressed as a weighted mix of the five channels:

Three tristimulus targets and five controls mean that several valid mixtures can produce the same chromaticity and brightness. That freedom is why the extra channels are useful. The solver can favour a mixture that uses less power, stays within channel limits, reduces spectral discontinuities, balances junction temperatures or more closely matches a reference illuminant.

A practical solver uses non-negative channel values and respects current limits. It also needs measured dimming curves because PWM duty, optical output and the human impression of brightness are three different things.

Chromaticity matching is not spectral matching

The XYZ matrix can place the beam at a specified white point, but CRI and TM-30 depend on the complete spectral power distribution. Two mathematically matching white points can render objects differently. A calibrated mixer therefore needs a second objective based on measured spectra. The alternative is to use carefully measured recipes rather than claim universal colourimetric accuracy.

Direct mode versus calibrated mode

Direct control remains useful for bench measurements and for selecting individual emitters. It should not be confused with fixture-to-fixture matching. Professional manufacturers measure and calibrate individual luminaires because LED wavelength and output vary by bin and temperature. Even nominally identical emitters are not perfectly identical.¹¹

A central finding of the project is that amber and phosphor-converted lime normally sit inside the triangle formed by red, green and blue on a CIE chromaticity diagram. They do not automatically add a new boundary colour. Their value is that they provide brighter and spectrally fuller ways to reach colours that RGB could already match.

The first model used the Cree XNPAPL package: red around 625 nm, green around 528 nm, royal-blue territory around 458 nm and a modelled PC-lime peak around 558 nm. For those primaries, adding lime left the chromaticity hull area unchanged to four decimal places. Yet the predicted brightness and colour-rendering results changed dramatically.¹³

The emitter choices are deliberately different

The first hardware is RGBL rather than the final five-colour fixture. A four-die Cree package places every source within one optical footprint, which simplifies colour mixing and makes it an appropriate first board. The RGBAL version uses discrete emitters and gains flexibility at the cost of more demanding mechanical and optical alignment.

Why the working white point became 3000 K

The project swept the Planckian locus from 2200 K to 6500 K using the modelled emitters. In that calculation, 3000 K gave the best overall balance of CRI Ra, TM-30 Rf and saturated-red rendering R9. The 2700 K recipe improved Ra by only 0.1 point but reduced R9 by almost 12 points. That is a poor trade for skin, timber and other warm materials.

Two findings of particular relevance to the video

Royal blue really can change the gamut. Unlike amber and lime, swapping blue for a shorter-wavelength royal blue moves an outer vertex of the RGB triangle. The project model predicted about 4.8% more chromaticity area, along with better TM-30 Rf and efficacy. Standard blue retained a small Ra advantage, which is a useful reminder that Ra and TM-30 do not ask the same question.

An optimised warm-white recipe may exclude the green die. At and below 3000 K, the model's best white switched the narrow green emitter off. Broad lime peaked in approximately the same region and provided the required luminance more efficiently. Green remains important for saturated greens and cyans, but it was not the preferred contributor to this particular white recipe.

The fork avoids teaching hundreds of effects about amber and lime. In the reported WLED 0.16-dev codebase, WLED applies its perceptual processing before the PWM bus sees the colour. The multi-primary bus then receives a linear-light RGB request, interpolates the calibrated table and produces four or five physical channel values.¹³

What the simulation says so far

For the modelled Cree RGBL fixture at 3000 K, the 17³ profile reported CRI 89.1, TM-30 Rf 87.8 and a 3005 K white. Of 4,913 grid points, 327, or 6.7%, needed gamut mapping. Across 2,898 in-gamut test colours passed through the quantised LUT, the reported mean ΔE 2000 was 0.02 and the maximum was 1.40.¹³

These low interpolation errors indicate that the lookup-table implementation adds little error relative to its underlying model. They do not demonstrate that the emitter model itself is accurate. Low numerical error within an inaccurate spectral model would still produce inaccurate output.

Controller

Use an ESP32 rather than an ESP8266 for this build. It provides adequate processing headroom, LittleFS capacity for the calibration profile and a substantially more camera-compatible default PWM frequency in WLED. Keep the controller thermally and electrically separated from the high-current LED stage where practical.

Output stage

WLED’s GPIOs are control signals, not LED power outputs. A constant-voltage strip needs a suitable switching device per channel, while high-power emitters need regulated constant-current drivers. Include defined off-state bias, gate or input resistors, hardware current limits, fusing and a safe power-up state. WLED’s own analogue-strip guidance also stresses external MOSFET circuitry and protection against overload and over-temperature.¹

Thermal design

Characterise the worst credible mixture, not merely one emitter at full output. A centre blend can activate amber and lime together, and a custom RGBAL recipe may operate several channels at high current. Place a temperature sensor near the LED substrate, apply controlled thermal derating and test after the heatsink reaches equilibrium. Colour shifts as the engine warms, so thermal stability is also colour stability.

Optics

Several visible source colours need a proper mixing distance, reflector or homogenising optic. The four-die Cree package gives the first RGBL build a shared optical footprint. The discrete RGBAL version will be less forgiving. A diffuser can hide the individual emitters when viewed head-on yet still cast coloured shadows onto nearby objects. Test beam uniformity at the actual installation distance, including the edge of the field and at low dim levels.

WLED’s documentation lists a default analogue PWM frequency of 19,531 Hz on ESP32, compared with 880 Hz on ESP8266.¹ This is an important reason to use ESP32. The frequency is high enough to avoid many common camera artefacts, but it does not provide a universal guarantee: shutter time, frame rate, rolling-shutter readout, PWM depth, driver behaviour and low-duty operation can still interact.

Higher PWM frequency also trades against PWM resolution and switching loss. WLED’s current bus code derives available bit depth from the PWM frequency and clock, so “as high as possible” is not automatically the best setting.⁴

The broader flicker literature is clear that frequency is only one variable. Modulation depth, duty cycle and waveform also matter, particularly when the light is being recorded by a camera. The US Department of Energy recommends testing across the dimming range and measuring the actual output rather than assuming a driver is flicker-free.¹²

1. Stabilise the mechanics, currents and temperature. Finalise the heatsink, optic, drive current and permitted simultaneous-channel load. Record cold-start drift and wait for thermal equilibrium before building the profile.
2. Drive one emitter at a known duty. The fork's raw override bypasses the LUT, auto-white handling and WLED's perceptual brightness curve. That bypass matters: without it, a nominal 50% request can be closer to 12% duty and the calibration would learn the wrong transfer function.
3. Capture every spectrum at several levels. The planned Avantes AvaSpec 2048-2 measures each die while duty is swept. Measuring the levels matters because driver droop, LED heating and supply sag can bend the flux-versus-duty curve.
4. Build the real channel model. Integrate each spectrum against the CIE colour-matching functions to obtain its XYZ contribution. Stack those results into the 3×N matrix and replace the provisional emitter model with the finished fixture's data.
5. Choose white point and headroom. Sweep the available luminance ceiling and choose the knee where more brightness starts costing a disproportionate amount of colour quality.
6. Solve the colour cube. The current toolchain solves 4,913 small convex problems for the 17³ grid. In-gamut points balance colour, rendering and efficiency. Out-of-gamut requests are reduced in chroma while hue and lightness are held as steady as possible.
7. Upload safely. The device checks the profile signature, version, geometry and CRC, then hot-loads it. A bad file is rejected and the previous calibration remains active.
8. Close the loop. Command known colours, re-measure the beam and report ΔE 2000 against intent. Repeat at low output, after thermal soak and across the expected ambient range.
9. Turn measurements into presets. Put the proven architectural and camera looks into WLED so the YouTube demo and final installation recall exactly the same states.

Why an 85% luminance ceiling can improve overall performance

At the absolute luminance ceiling, the three colour constraints can completely determine the channel mix. The optimiser has no freedom left to improve the spectrum. Backing away from that ceiling gives it room to choose a better mixture.

In the model, giving up 15% of peak luminance bought 8.9 CRI points, nearly 38 R9 points and about 5% better efficacy. WLED's master brightness then scales all channels together, so the spectral mix chosen at the top of the curve remains relevant at every lower brightness.¹³

What if there is no spectrometer?

A tristimulus colour meter can provide useful xy/uv and luminance data for the basic mixing matrix. It cannot reveal spectral discontinuities or calculate reliable rendering metrics without spectral information. A camera or phone is useful for demonstrating the result, but it is not suitable as the only calibration instrument. For one comprehensive calibration session, borrowing or hiring a spectrometer is likely to provide more value than another iteration of unmeasured firmware.

What can still go wrong on the real fixture

• Ambiguous naming: RGBWW does not tell us whether there are four or five controllable channels.
• Common-anode versus common-cathode hardware: the wrong driver topology or polarity setting produces more than a software problem.
• Auto-white surprises: WLED can derive white from RGB. In the repurposed RGBAL mapping, that means it may add amber or lime unexpectedly. Use manual white during characterisation.
• CCT blend heat: additive centre blending can run both secondary channels hard. Current-limit and thermally test that case.
• Power estimation: WLED’s brightness limiter is useful, but a custom high-power analogue engine still needs hardware limits and a supply sized from measured loads.
• Thermal colour drift: flux and wavelength move with temperature. A temperature-stable fixture is easier to calibrate and easier to match.
• Optical separation: a diffuser that appears uniform at ten centimetres can still cast separated colour shadows at the wall.
• Low-level dimming: the bottom few per cent expose PWM resolution, driver turn-on thresholds, channel mismatch and camera banding.
• Calibration by eye: our visual system adapts quickly. Side-by-side references and instruments prevent “looks about right” from becoming a permanent coefficient table.
• Network exposure: WLED is a convenient local controller. Do not expose a luminaire’s web interface directly to the public internet.

RGB supplies the edges. Amber gives warmth without a red-green compromise. Lime fills a useful, efficient part of the middle spectrum. Together they create a flexible platform for an architectural light that can do bold colour and far more nuanced pale light than a basic RGB engine.

WLED is a suitable foundation because it already provides Wi-Fi provisioning, a developed interface, presets, APIs and effects. Stock WLED's RGBCCT path is useful during initial hardware commissioning. The custom fork then adds the required fixture-specific conversion from ordinary colour requests to measured RGBL or RGBAL drive values.

The supplied project results make a strong engineering case, but they are still model results. The first Cree XNPAPL RGBL board, its real lime spectrum, PWM waveforms, optical mixing and closed-loop ΔE measurements will decide whether the model survives contact with the bench.

It also gives the YouTube video a clear structure. Start with the familiar WLED control system, then explain why five physical outputs are not the same as five-primary colour management. The key demonstrations are the unchanged gamut triangle, the green channel switching off in warm white, and the improvement in colour quality gained by reducing peak luminance by 15%.

Together, these results show how an experimental colour fixture can be developed into a repeatable architectural luminaire.