Why the lighting is the hard part.

Patient self-capture by smartphone. Mobile-friendly photo apps positioned as "good enough." Low-cost lightbox clones that promise the same dataset for a tenth of the price. They all fail at the same step, and the failure has a name a century old.

The shortcuts that have been tried — and why they didn't work.

Before pink, the trial team explored what every grant-funded medical-imaging effort considers: substituting dedicated hardware with cheaper, lighter, faster alternatives. Several were attempted in parallel by partners under the consortium umbrella, by independent research groups, and by would-be vendors looking to bid against PhotoRobot on cost rather than capability.

Three categories emerged. None survived contact with the data.

1. Patient self-capture via consumer smartphone apps

The most attractive approach in theory: patients photograph themselves at home, upload to a clinician-managed app, and the AI proceeds from there. No clinical time required, no waiting-room logistics, no per-patient hardware cost.

In practice, the inputs varied across every controllable axis at once. Lighting (overhead vs. window vs. lamp), distance (arm-length vs. set-and-step-back), framing (waist-up vs. shoulder-up), focal length (wide-angle distortion from selfie cameras vs. dedicated rear camera), colour temperature (incandescent vs. daylight vs. fluorescent), exposure (auto-bright vs. manually-tapped), background (cluttered bathroom vs. plain wall), and pose discipline (consistent stance vs. casual). Each patient's session looked different from the next; their own follow-up sessions looked different from their baseline.

AI training on this kind of variability does not produce predictions. It produces noise wearing the costume of predictions — a model confidently outputting unrelated results that look authoritative until measured against ground truth.

2. Low-cost lightbox clones

The category that arrives every time medical imaging gets EU funding attention: small enclosed devices, often built around a single LED panel and a fixed camera, marketed as "the same thing as pink for €15,000." Hardware-wise they look serious. The shells are clean, the cameras are mid-range mirrorless, the device has a startup screen with a logo.

The trade-off is invariably hidden in the same place — the lighting. Studio-grade flash or continuous lighting is the largest BOM line in any serious photographic device. Cut it, and the cabinet still closes; the device still produces images; the price point still hits the marketing target. But the images are unstable across operating conditions (line voltage flicker, ambient temperature, panel aging), and unrepeatable across units (each LED panel coming off the line varies in CRI and colour temperature by a margin that matters for training data).

The pattern is consistent enough to predict: small device with perfectly mediocre output, sold once, never repeats. The grant gets spent, something gets demonstrated, the institutional partner moves on. The original consortium loses time it cannot recover.

3. Operator-driven manual studio shoots

The variant that looks competent on the surface: a clinical photographer is assigned to the trial, given a Canon body and a softbox, and asked to standardise the protocol. The first month of output is excellent — because the photographer is engaged, the protocol is fresh, and the institution is paying attention.

Then the photographer goes on leave. A different one steps in. Cross-site coordination drifts. The softbox gets repositioned to make room for an unrelated equipment delivery. Hospital photo-room renovation moves the studio across the building. The protocol document gets revised; the revision doesn't propagate; some sites are on v3, some on v4. Six months in, the dataset is no longer a dataset — it is a collection of operator personalities, none of which the AI can absorb cleanly.

Why lighting is the single hardest part.

The reason all three shortcut categories fail traces back to one component. Lighting is the part of a photographic system that most determines output consistency, and the part that is most expensive to do correctly.

Professional studio lighting — the kind PhotoRobot has used since the company started — comes from a small group of suppliers whose product lines exist precisely because the requirement does not have cheap solutions. The brand list, depending on configuration:

These are not commodity items. They are colour-accurate, voltage-stable, flicker-free, calibrated to fall within tight CRI tolerances, and engineered for replicable behaviour across deployed units. Two heads of the same model, built two years apart, will produce indistinguishable output. That is the requirement; everything else is a workaround.

What pink locks down by design.

Beyond lighting, several other variables have to be controlled or the multi-site dataset breaks. pink enforces each one — hardware-wise where possible, software-wise where not:

• Camera body and lens. Canon optics, fixed configuration, identical across deployments.
• Working distance to subject. Built into the cabin geometry — the patient cannot stand too close or too far, by physical constraint.
• Background reflectance and colour. Cabin interior surfaces are spec'd; replacement panels match.
• Patient positioning and pose protocol. Pose markers are physical, not "remember to stand here" — the patient's feet have a place.
• Pose sequence. Capture order is fixed by the trial protocol and runs automatically; no operator decision required.
• Metadata schema. Each capture stamps timestamp, anonymised ID, pose code, machine serial, exposure parameters. No post-hoc reconstruction.
• Anonymisation pipeline. Identifying detail strips before any image leaves the device; the unanonymised version does not get a network hop.

Cross-site identical machines produce cross-site identical capture conditions. Cross-site identical capture conditions produce a dataset the AI can actually train on. That is the value proposition of dedicated hardware in a multi-centre clinical trial — the value that gets routinely overlooked when the temptation arises to cut costs with consumer alternatives, and is paid for in extended timelines and dataset rework when the cost-cutting attempt fails.