The Studio Lighting Platform

A room with two jobs

We have lit two of the studio's three spaces already, and both went well by the ordinary standard: the light is good and nobody thinks about it. The downstairs gallery is the first one where that standard is not enough, because this time the lighting is also the thing we want to build software against.

Those two jobs pull against each other, and understanding why is most of the work. Architectural lighting is designed to be installed once and then forgotten. It optimises for a clean ceiling, a long life, and controls simple enough that a visitor can use them. Entertainment lighting is designed to be reprogrammed nightly. It optimises for precision, speed and individual control of every fixture, and it looks like rigging because in a venue nobody is looking at it.

We want the first one's appearance and the second one's capabilities. That combination exists, but it sits in a narrow part of the market, and the products that appear to offer it without the narrowness — Hue, Casambi, Tuya — are the ones that stop dead at a ceiling you cannot pass. The trouble with those ceilings is that you find them after you have bought the fixtures.

This document sets out what we found, what each technology is actually for, how the whole thing fits the systems we already run, what it costs across four clearly-labelled price brackets, and — most importantly — the order in which we should spend the money so that the risky decision gets tested for a couple of hundred euros before it gets committed for several thousand.

Two lineages of lighting control

Almost every lighting control protocol descends from one of two families, and knowing which one you are holding explains most of its behaviour, including the parts that seem arbitrary.

The entertainment lineage begins with DMX512, designed in the 1980s for theatre. A controller broadcasts a stream of 512 channel values, many times a second, down a cable. Fixtures listen and obey. Nothing is acknowledged, nothing is stored, nothing is asked. It is deliberately dumb, and the dumbness is what makes it fast — the whole universe of 512 channels refreshes up to 44 times a second whether one fixture is listening or two hundred are.

Three modern descendants matter to us. sACN (formally ANSI E1.31) carries the same data over an ordinary ethernet network using multicast, which means a managed switch can deliver each stream only to the devices that asked for it. Art-Net does much the same thing but older and by broadcast; it matters because vvvv speaks it natively. DDP is a leaner variant used for very high pixel counts, where the overhead of DMX's universe structure starts to hurt. RDM is the one genuinely two-way addition: it lets a controller discover fixtures, set their addresses and read their status remotely, instead of someone standing on a ladder with a screwdriver.

The architectural lineage begins with DALI, and its priorities are the opposite. A controller addresses one specific fixture, sends a command, and waits for it to land. Fixtures store their own scenes and report their own faults. It runs over any two-core cable, tolerates bad wiring, survives decades, and is a genuinely good standard — open, multi-vendor, and the thing most commercial buildings are wired for. It is also, by design, slow.

Around DALI sit the wireless variations. Zigbee is an open low-power mesh; used through the Hue bridge it is a closed product, but used through Zigbee2MQTT the same bulbs become fully local and scriptable. Bluetooth mesh, as commercialised by Casambi, is the most elegant of the consumer-adjacent options and genuinely beautiful to install — and its API tiers are where our kind of experimentation runs out of road. KNX is the building-automation heavyweight: open, capable, expensive, and slower than DALI for anything visual. And 0–10 V is the oldest survivor, a single analogue voltage that dims a driver — utterly reliable and completely unaddressable, so one wire dims one group and that is all it will ever do.

The speed split

The central finding is that the choice between those two families is decided by physics rather than preference, and it cannot be worked around in software. It is worth being precise, because the size of the gap is what makes the decision straightforward.

A DALI command frame takes 15.83 milliseconds to transmit, and the standard requires a further 13.5 to 21.2 milliseconds of settling before the next one. One addressed command therefore costs about thirty milliseconds. Addressing twenty fixtures individually takes roughly two thirds of a second — call it one and a half refreshes per second. Colour changes are worse again, because each one needs five or six frames per fixture, which puts twenty fixtures at well under one refresh per second.

DMX sends all 512 channels in a single frame up to 44 times a second, and adding fixtures does not slow it down at all, because every fixture reads the same broadcast.

"Thirty times faster, and unlike DALI it does not get slower as you add lights. That is not a tuning difference, it is a different category of thing."

None of which makes DALI bad. It makes DALI a scene-recall protocol, and scene recall is most of what architectural lighting is asked to do — hold a room at a comfortable level, fade gently between presets, come on when someone walks in. It simply cannot produce movement that the eye reads as movement, and a room built as a lighting-software testbed whose fixtures refresh once a second is not a testbed.

Speed is only half the picture, though. The other half is how far a technology lets you push before you hit a wall, which is the thing we have been burned by before. Plotting both axes together sorts the whole landscape rather cleanly.

The quadrant makes the recommendation for us rather than the other way round. The top-right region — fast enough for effects, open enough to never be blocked — contains the DMX family and essentially nothing else. Everything on the left is architecturally excellent and too slow to be the experimental layer. Everything at the bottom is where we have been burned before.

So the architecture does not try to unify the two families at the wire. It unifies at the control plane, where everything is addressable from one place, and accepts two physically distinct transports underneath: DMX for anything that has to move, DALI and Zigbee for anything that simply has to be on.

What the studio already runs

The reason this project is smaller than it looks is that most of the software already exists and is already running. It is worth listing, because each piece has a job in the lighting platform and none of them needs to be bought or learned.

Node-RED and Mosquitto on connect.spacemusic.tv give us an MQTT broker and a visual patchbay for wiring things to other things. MQTT is the natural boundary between the lighting world and everything else — it is how a sensor reading, a button press or a scene change crosses from one system to another without either side needing to know much about the other.

Centrifugo on relay.spacemusic.tv already pushes live data to browsers in real time. Pointing it at fixtures instead of pixels on a screen is a small step, and it is the piece that makes "drive the lighting from our own data" a weekend project rather than an integration.

Grafana gives us dashboards over lighting telemetry alongside everything else we monitor, and Authentik gates any web control surface we build, exactly as it already gates the prototypes host.

vvvv gamma is the piece that is genuinely unusual, and it is the reason this project is worth doing at all. It is a live programming environment where the program runs continuously while you edit it, built for real-time visual work, and it has first-class support for lighting protocols. Most studios attempting this would be writing a show engine from scratch. We would be opening a patch.

The one addition is Home Assistant, running locally on-site. It is worth adding rather than building around, because it brings Zigbee support, presence handling, schedules, phone and tablet interfaces, and ready-made DALI gateway integrations — several months of work that we would otherwise reimplement worse. The constraint is that it stays out of the real-time path, and it must run on hardware in the building so that the gallery keeps working when the Hetzner box does not.

Where lighting joins the estate

Put together, the shape is two columns under a shared boundary. The estate we already run sits at the top. MQTT is the seam. Below it, vvvv owns everything fast and Home Assistant owns everything stateful, and they touch only through that seam.

The rule that keeps this honest is the one written on the bottom of the diagram. Home Assistant is a stateful platform that restarts when it updates, and a restart must never be able to stutter a running show. Keeping it strictly out of the real-time path costs nothing and removes an entire class of failure that would otherwise be very hard to diagnose.

Six conductors and an open question

Three-phase track — the kind we already use — carries five conductors: three independently switchable mains circuits, a neutral, and earth. What makes some track interesting is a sixth and seventh: a dedicated pair of small data conductors running the full length of the profile, which any fixture can tap anywhere along the run.

There is a fork here that is easy to walk into blind, because both branches are sold as "DALI track". One design — ERCO's, and one of Zumtobel's — repurposes two of the three phases as the data pair. Same profile, same appearance, but the track drops from three 16 A circuits to one, and the pair is locked to DALI permanently. The other design adds a separate control pair, keeps all three circuits, and its conductors are protocol-agnostic.

Stucchi's ONETRACK — the current form of EUTRAC, which Stucchi acquired in 2021 — is the second kind. Its own specification reads L1/L2/L3/N/PE, 16A/440V + D+/D- 2x1A/50V FELV AC: six copper conductors plus earth through the extruded body. It costs around €30 a metre, the adapter that taps the data pair costs about €13, and a Swiss distributor in Schlieren stocks it. The infrastructure is cheap. Effectively all of the money is in the fixtures.

The plan is to run DMX rather than DALI on that pair. And here is the part that has to be said plainly rather than buried: that is not yet proven for our market.

The evidence in favour is real. Stucchi's own installation sheet describes the pair as carrying "a general data signal (for example dali, dmx, lon, eib)". ETC — an American manufacturer with about as much DMX authority as exists — sells this same extrusion in North America explicitly specified for DMX and RDM, rated at 32 devices per run, with a matching terminator. PXM in Poland ships track spotlights that pick DMX off exactly these conductors.

The evidence against is also real. Every current European document labels the bus DALI; Stucchi's 2024 brochure mentions DMX zero times, and their site's product search returns nothing for it. ETC's version will not ship to Switzerland and runs on a different voltage. No European DMX terminator for this track is catalogued by anyone. And underneath the paperwork sits a physical question: DMX is RS-485 at 250 kbit/s and expects shielded twisted pair, whereas this is two parallel bare busbars, untwisted and unshielded, millimetres from 400 V mains, with a stub at every adapter and a reflection at every joint. DALI runs at 1 200 bit/s and does not care. DMX is two hundred times faster and might.

Which is precisely why the next thing we buy is not fixtures. More on that in §9.

The layers beyond the spots

Track spots are only one of three layers, and the other two are where a lot of the interesting behaviour will actually come from.

Pixel strip — where colour belongs

Colour-changing spotlights are expensive and, more importantly, they are usually a colour-quality compromise: mixing narrow-band red, green and blue emitters renders objects worse than a good white LED does. Addressable strip has the opposite economics. A metre of strip has dozens of individually controllable pixels, costs very little, and lives in a surface aluminium profile in a cove or shadow gap where nobody looks at it directly.

The controller layer here is well-solved and cheap. An ESP32 running WLED costs about ten francs, speaks sACN, Art-Net and DDP natively, and is completely hackable. For runs that must always work — the ones that would be embarrassing if they glitched during a viewing — a proper pixel controller such as an Advatek PixLite is the reliable spine, DIN-rail mountable and accepting both protocols at once. Starting on WLED and graduating specific runs to a PixLite as they earn it is the sensible progression, and both speak the same transport so nothing has to be rewritten.

One practical note that saves pain later: buy controllers that accept sACN and Art-Net simultaneously. It costs nothing and removes a whole category of future regret, because vvvv speaks Art-Net natively while sACN is the better choice for the permanent install.

Sensing — position, not presence

The obvious sensor is a motion detector that turns lights on when someone walks in. The more interesting one is mmWave radar, and specifically the LD2450, which has an official ESPHome component and tracks up to three people's positions — distance, angle and speed — rather than emitting a single presence bit.

That difference matters for a gallery. A presence sensor lets the room respond to whether someone is there. A position sensor lets the room respond to where they are standing and which way they are moving — light that follows a visitor along a wall, or brightens the piece they have stopped in front of. It is a genuinely novel thing to build against, the sensors cost about twenty francs, and it is the single cheapest source of interesting behaviour in the whole project.

The software path out of vvvv

One concrete finding worth recording, because it would otherwise cost someone an afternoon. There is a community sACN library for vvvv, VL.sACN, and it works — but it hardcodes the sACN priority value and has no sync support, which rules it out for the layered-priority work we would want to test. The better path is Haukcode.sACN, an ordinary .NET package with a settable priority, real sync packets and automatic multicast addressing. vvvv can reference arbitrary .NET libraries, so it runs in-process with no extra hardware in the signal path.

Fixtures, by price bracket

Before the numbers, one finding that widens every option below. The price of a fixture is not the price of its colour quality. Verified European retail sells CRI 97 three-phase track spots at €180 and CRI 90 at €44 — so the premium for the top of the colour-rendering range is roughly €30 to €60 on a 30 W head. On a €500 fixture, the other €400 is control electronics, tunable white, optics and warranty.

And the bar itself is lower than instinct suggests. Getty and the Canadian Conservation Institute both put art viewing at CRI above 90, describe CRI as "a target, not a firm criteria", and treat R9 — the deep-red score — as the number that actually matters, with R9 above 90 counted excellent. A CRI 90 fixture with strong R9 is entirely defensible for hanging work.

The structural finding is less comfortable: there is essentially no inexpensive DMX-native architectural track spot on the European market. Cheap high-CRI track spots exist in quantity, and they are all DALI, Triac or 0–10 V. DMX-native gallery heads start around €310. That gap is what the brackets below navigate.

All figures are indicative European retail gathered in August 2026, mostly including German or Italian VAT. None of them is a Swiss landed cost — export removes German VAT but adds Swiss import VAT, customs clearance and, for long track, a bulky-goods surcharge. Treat every number as a starting point for a quote.

Two constraints cut across all four brackets. Optics are a mandatory line item on almost every gallery head and are quoted separately by every vendor — assume €85–150 for a wash zoom and up to €190 for a profile lens. And in brackets A and B, where we assemble a CE-marked head with a CE-marked driver, the CE marks do not transfer to the combination; in a fixed Swiss installation the liability sits with whoever connects it. That question needs asking early, because a no from the electrician removes both of the cheaper brackets.

Testing small before committing

Everything above rests on one unproven assumption, and the correct response to an unproven assumption that costs four figures to get wrong is to spend three figures finding out.

The staging below is ordered by commitment rather than by time. Each stage costs more and is harder to reverse than the one before it, and the first one is deliberately tiny.

Stage 0 — the bench rig

A desk, not a room. Two metres of ONETRACK offcut, an end feed, two data adapters, one cheap head, one DMX/RDM driver, a 120 Ω terminating resistor, and an isolated USB-DMX interface. Around €150–250 in total.

The test that matters is deliberately unglamorous: run DMX over the busbars at the full 44 Hz, with the mains loaded on all three phases, in a strictly linear chain with the far end terminated, and leave it running for hours with an analyser watching the error rate. Minutes prove nothing — the failure mode is intermittent corruption, which is exactly the thing that would later be impossible to distinguish from our own bugs. While the rig is on the desk it also answers the lens question, proves the vvvv and Home Assistant chain, and lets us walk past an LD2450 to see whether position-tracking is as interesting in practice as it is on paper.

If it passes, the architecture is confirmed and Stage 2 can be specified with confidence. If it fails, we pull a shielded twisted pair alongside the track and use the busbars for power only. That costs cable and nothing else — provided we find out now rather than after the fixtures are up.

Stage 1 — the control plane, with zero new fixtures

The switch and lighting VLAN, Home Assistant on a small machine in the building, the MQTT bridge to the existing broker, and a DALI gateway onto the office spots that are already installed and already capable and currently sitting on ordinary dimmers.

This stage is worth calling out because it buys no gallery hardware at all and still produces something real: the office lighting becomes controllable from a phone and from a vvvv patch, on exactly the control plane the gallery will later use. If something about the approach is wrong, we find out here, upstairs, for the price of a gateway.

Stages 2 to 4 — and only then

Track and fixtures. Then the pixel and sensing layers. Then studio data driving the room, and a control surface behind Authentik. Each of those is straightforward once the two questions below it have been answered, and each is a reasonable place to stop for a while.

What we are deliberately not doing

Recording the rejections matters as much as the choices, because in a year someone will reasonably ask why we did not just use Casambi.

1. Casambi. Genuinely the most elegant minimal option on the market and a pleasure to install — and the same closed ceiling as Hue with better industrial design. Bluetooth mesh with gated API tiers. rejected on: extensibility
2. KNX. The building-automation standard: open, capable, and genuinely good at what it does. Expensive, slow for anything visual, and heavily over-engineered for one room in a studio. rejected on: cost and speed
3. Phase-sacrificing DALI track. Throws away two thirds of the power capacity and permanently locks the data pair to a protocol that cannot do effects. Two losses for no gain. rejected on: capacity and lock-in
4. 48 V magnetic track for the gallery. Visually lovely and genuinely reconfigurable, but no system exists that is simultaneously CRI ≥ 95 and wired per-fixture addressable — the industry answers per-module control with Casambi or Zigbee. Still fine for non-gallery accent runs. rejected on: no addressable high-CRI option
5. Hue as a platform. The bulbs are excellent and, through Zigbee2MQTT, entirely usable as local accents. The bridge is the part we decline. partially adopted
6. Direct import of unbranded fixtures. CRI 97 Citizen-COB track spots are available at $28–32 from Chinese manufacturers, and some may well be fine. But the CRI claims arrive with no test report, no R9 figure and no bin code; the "DMX" claims sit on a marketing bullet next to five other protocols; the supplied adapters cannot reach the track's data conductors; and as a direct importer we have nobody to hold to the declared-CRI tolerance that Swiss regulation would otherwise give us. available, characterised, not recommended

Why this matters

"The expensive mistake is not buying the wrong fixture. It is buying a control path that cannot be extended, and only finding out two years later."

Everything in this document resolves to one decision with a long shadow. Fixtures get replaced — that is normal, and track exists precisely to make it easy. Control paths do not get replaced, because replacing one means opening the ceiling again.

If the gallery's data conductors carry DALI, then in five years the room will recall scenes slightly better than it does today and nothing else. If they carry DMX, the room can do anything we can write, at frame rate, for as long as we care to keep writing it — driven by our own data, through our own relay, from a patch in an environment we already know.

That is the whole argument for taking the harder path here. Not because DMX is better in the abstract — for most architectural work it plainly is not, which is why the industry settled on DALI — but because we are the unusual case that will actually use the difference. We have the software estate. We have the vvvv skills. We have a room whose explicit purpose is to be pointed at.

And the immediate next steps are small, cheap and reversible: build a bench rig for the price of a nice dinner, and send two emails asking for quotes. Both can happen before anything is committed. Everything else in this document sits downstream of one measurement, taken on a two-metre offcut of aluminium, on a desk.

Glossary

Terms and acronyms used in this document, in plain language.

Art-Net

An older way of carrying DMX over an ethernet network. Broadcast-based, very widely supported, and what vvvv speaks natively.

Casambi

A commercial Bluetooth-mesh lighting system. Elegant and easy to install; its programmability is gated behind API tiers.

Centrifugo

The real-time message relay we already run at relay.spacemusic.tv, used to push live data to browsers.

CRI

Colour Rendering Index. How faithfully a light shows colours, 0–100. R9 is the specific deep-red score and matters more than the headline number for artwork.

DALI

Digital Addressable Lighting Interface. The architectural standard — two-wire, robust, addressable, open, and deliberately slow at roughly 1.5 updates per second across twenty fixtures.

databus

The extra pair of small data conductors running the length of some three-phase track, tappable anywhere along the run by a suitable adapter.

DDP

Distributed Display Protocol. A leaner alternative to sACN used when pixel counts get high enough that DMX's universe structure becomes overhead.

DMX512

The entertainment lighting standard. A controller streams 512 channel values up to 44 times a second; fixtures listen and obey. Fast, stateless, unacknowledged.

DT8

The DALI extension handling colour and colour temperature. Several times slower again than plain DALI dimming.

ESPHome

Firmware that turns an ESP32 microcontroller into a Home Assistant device from a short YAML description, without writing embedded code.

FELV

Functional Extra-Low Voltage. Low voltage whose shock protection still depends on the mains circuit behind it — so it must be treated as mains-adjacent wiring, unlike properly isolated SELV.

Home Assistant

An open-source home and building automation platform. Brings Zigbee, presence, schedules, dashboards and DALI integrations for free.

LD2450

A 24 GHz millimetre-wave radar module that tracks up to three people's positions rather than just detecting presence. Officially supported by ESPHome.

MQTT

A lightweight publish-and-subscribe messaging protocol. Our seam between the lighting world and everything else; we already run a broker.

ONETRACK

A.A.G. Stucchi's current three-phase track system, and the continuation of the EUTRAC line Stucchi acquired in 2021.

PWM

Pulse-Width Modulation. How LED drivers dim, by switching very fast. A low switching frequency causes banding on camera, which is why selectable PWM appears in fixture specifications.

RDM

Remote Device Management. The extension that lets DMX talk back — discovering fixtures, setting addresses and reading status remotely rather than by hand at a ladder.

RS-485

The electrical signalling standard DMX512 rides on. It expects a shielded, impedance-matched twisted pair — which is exactly what track busbars are not.

sACN

Streaming ACN, formally ANSI E1.31. The modern way of carrying DMX over a network, using multicast so switches deliver only what each device asked for.

three-phase track

Standard architectural track carrying three independently switchable mains circuits plus neutral and earth. Fixtures clip in anywhere and move freely.

vvvv gamma

The visual live-programming environment the studio uses for real-time work. Programs keep running while you edit them.

WLED

Open-source firmware turning a ~CHF 10 ESP32 into an addressable-LED controller that speaks sACN, Art-Net and DDP.

Zigbee2MQTT

Software that talks to Zigbee devices directly and republishes them over MQTT — the route to using Hue bulbs locally without the Hue bridge.