The hypostyle hall at Karnak is a forest of 134 sandstone columns covering roughly half a hectare, with the central twelve at the nave reaching about 21 metres. It is one of the most heavily photographed buildings in the world. It is also one of the hardest to see — the columns are dense, the visitor paths are narrow, the lighting in the hall has been engineered for a century around the constraints of mass tourism, and the polychromy that once covered every surface is now visible only in fragments at the column capitals.

The VR reconstruction we have been building since 2024 is an attempt to give the hall back the conditions under which it was designed to be experienced. This is a field note on what the model does, what it does not do, and the one decision that has consumed more of our time than the entire geometry pipeline.

The geometry was the easy part

The photogrammetric source set is 47,184 images: a combination of our own DSLR work in 2024–25 (a Sony A7R V at 60 megapixels, mostly on monopod), the CIPA-RecorDIM open archive, and a 2018 drone survey by the Polish Centre of Mediterranean Archaeology that has been released for academic use. Aligning that many images is not theoretically interesting any more — Agisoft Metashape on a single workstation with 256 GB of RAM did the bundle adjustment in eleven days, mostly unattended. The dense cloud was 4.1 billion points. The decimated mesh that ships in the VR build is 81 million triangles after aggressive but careful simplification with InstantMeshes.

The geometric accuracy of the final mesh, measured against 38 control points surveyed in 2025 with a Leica TS16 total station, is ±3.8 mm RMS. That is more than enough to render the hall to within the limit of what any headset display can resolve at conversational distance.

The hard part is the light

The hypostyle hall was originally lit by clerestory windows along the central nave, oriented to admit direct sunlight only for a few hours per day. Most of the time the hall was dim — not dark, but lit by reflected daylight, by oil lamps in the side aisles, and during festivals by the great processional torches that we know about from the Opet relief sequences on the western side.

You cannot photograph that light. It is gone. We have a very good model of the building, and a fair model of the sun (we know the latitude, we know the precession-corrected solar geometry for the Ramessid period), but the air in the hall — the dust, the smoke from the lamps, the haze from the burnt offerings — we have to reconstruct from period descriptions and from analogue measurements in still-functioning lampblack-lit interiors. We have spent six months on it and we are not done.

What we have shipped, for now, is a single calibrated lighting state: high noon at the spring equinox, year 1252 BCE, clear sky, no lamps lit. This is the state most stable across our sources and the one we are most willing to defend.

Geometry is a measurement. Light is a hypothesis. Be very careful to mark which is which when you put a reconstruction in front of a non-specialist audience.

What the model deliberately does not do

Three things, all of them deliberate.

It does not show polychromy on surfaces where the Ramessid pigment has not been securely identified. The capitals carry the colour scheme documented by the Centre franco-égyptien d’étude des temples de Karnak survey teams. The lower column shafts, where the surviving pigment is heavily contested, are rendered in plain stone with a small caption marking the omission. We will not paint a surface for visual richness if we cannot cite the source.

It does not insert human figures, animals, or atmospheric effects beyond the calibrated lighting state. Adding a single statue, a single offering table, a single priest, immediately turns the reconstruction into a diorama, and once you have a diorama you are inviting the viewer to read it as a documentary record of a specific moment. We are not equipped to do that.

It does not run the full mesh on a Quest 2 or earlier hardware. We tried. The frame budget collapses around the central twelve columns and the comfort cost on a sustained walkthrough is unacceptable. The Quest 3, the PCVR builds, and the Apple Vision Pro hand off the rendering load well enough; we are not going to ship a degraded version of the mesh just to widen the hardware support.

What it is useful for

Three audiences so far, in order of how much we have learned from them.

Egyptologists working on the hall — particularly on the procession routes and the visual axes — have used the VR build to test sightline arguments that you cannot easily test in the physical hall because the visitor paths are fixed. Two papers under review at the moment use the model in this way and one of them is going to revise an axis-of-vision argument that has been in the literature since 1973.

Conservators at the Karnak site have used the model as a baseline for condition monitoring. The 2025 mesh is the reference against which the 2026 re-scan will be differenced, and the difference at 4 mm resolution is enough to surface column-base spalling that is otherwise easy to miss against the variation in the sandstone.

Visitors to the Australian Museum, where the public-facing VR booth has been running since February, have largely told us that the hall feels bigger in the model than in real life. The reason is almost certainly that the model removes the modern fences, the staging, the railings, and the tourists. Whether that is a feature or a problem is a question I keep asking myself and do not yet have a stable answer to.

What we are doing next

The next build, due in October, will add the calibrated late-afternoon lighting state (5:40 pm at the autumnal equinox — the moment when the western light reaches the deepest into the hall). We will also start to layer in the documented Opet festival route, at the explicit request of the team at CFEETK, but as a separate mode of the build with clear annotation, not as a default state of the model.

If you teach a course on the Ramessid period and would like a build for classroom use, write to [email protected]. The build is available without licence fee for non-commercial educational use; we ask only that you cite the model in any paper that uses it for an argument and that you do not redistribute the binary.

— Elara