A Scenario exploration · Ancient Egypt, restored

The Stones Remember

Five places along the Nile, each restored to one moment in its history and narrated by a guide. Walk inside them, then see the same stones as they stand today. Every panorama, 3D scene, voice and score was made with Scenario, generated through Scenario MCP on top of published surveys and Blender geometry.

Download PDF Download skill
  1. IThe Great Pyramid of KhufuGiza
  2. IIThe Great Hypostyle Hall, KarnakThebes
  3. IIIThe Tomb of TutankhamunValley of the Kings
  4. IVThe Temple of Isis, PhilaeAswan
  5. VThe Great Temple of Abu SimbelNubia

Choose a site in the list or on the map.

How it was made

The Stones Remember

The harbour below the Giza plateau at dawn, around 2560 BC: a painted stop view
The harbour at the foot of the Giza plateau at dawn, around 2560 BC: the painted view of the first Khufu stop.

Five narrated visits of ancient Egypt, each restored to one moment in its history. Every panorama, 3D scene, voice and score was made with Scenario, generated through Scenario MCP on top of published surveys and Blender geometry.

Sites
5
Stops
32
Scenario models
14
Credits
36,482

01 · The idea

From one room to five places

The request: an extremely accurate visit of Philae, Karnak, Abu Simbel, the Great Pyramid and the tomb of Tutankhamun, at the level of a professional VR production for museum education, with a guide, people in the scenes, voiceover and 360 or interactive 3D. The sequel to The Room Remembers, which showed one room across time, keeps its hook: every site pairs its historical moment with the same viewpoint today.

Why 3D here and 360 there

Khufu and Karnak are geometric, measured and repetitive (courses, corbels, a grid of 134 columns): scripted Blender geometry gives exact, walkable architecture. Abu Simbel, Philae and KV62 are dominated by sculpture, dense painted relief and objects, where 8192 px panoramas carry far more detail than realtime meshes.

The hub's Nile map: a Python base drawn from Natural Earth data, restyled by GPT Image 2.5 Sunburst, with the seven ancient Delta branches.

02 · Research and accuracy

Every pixel is a claim about history

No site agent generated an image before writing its research bible: dimensions, layout, materials, colours and dated features, each tagged KNOWN, RECONSTRUCTED or UNCERTAIN, with sources. Wikipedia counted only as a pointer. A second file, ACCURACY.md, maps every narrated sentence and every scene element to its source, and the narration says uncertainty aloud: “Egyptologists think”, “no one knows whether”.

  • Khufu Petrie 1883, Lehner and Hawass 2017, Tallet 2017 (Merer's papyri), AERA
  • Karnak Great Hypostyle Hall Project (Brand), Porter and Moss plans
  • Abu Simbel Porter and Moss VII, Burckhardt 1819, Belzoni 1820, Shaltout and Belmonte 2005
  • Philae Haeny 1985, Kockelmann 2012, Porter and Moss VI, Baedeker 1914
  • KV62 Griffith Institute journals and Burton photographs, Carter and Mace 1923, Reeves 1990
The Great Hypostyle Hall: 134 columns in 16 rows, set out from P. J. Brand's published plan and Porter and Moss.

03 · The 360 sites

A measured guide under every panorama

Image models draw plausible temples, not correct ones. Each 360 stop therefore started as a true-scale equirectangular blockout from the bible's numbers: the Abu Simbel hall at 18 by 16.7 m with eight 1.7 m pillars, the facade at 38 by 33 m with four 21 m colossi. The shared Blender render queue stayed blocked by other sessions all night, so the agents wrote small numpy ray-casters that render these guides on the CPU, with no Blender process and no GPU.

The guide then goes to the image model as structure reference, together with photographs of the real site where useful, and a yaw-by-yaw prompt that says what stands in each direction.

Abu Simbel, the Great Hall: the measured guide, eight 1.7 m pillars at true positions.
The painted panorama, around 1255 BC.
Tutankhamun, the sealed doorway: the guide from the tomb survey.
The painted panorama.
Philae, the southern landing: the guide from a sourced plan.
The painted panorama, around AD 100.

Before painting a site, each agent asked Scenario MCP to recommend models for the job at quality priority, searched the catalogue newest first, read each live schema, then ran a shoot-out on one node with the same inputs. The winner differed by site.

Edit models paint a rectangle, so a seam appears at yaw 180. The fix: roll the panorama half a turn so the seam sits in the centre, repaint it, then merge back only a feathered central strip after phase-correlation alignment. Perspective patches do the same for smaller defects.

04 · Then and Today

The same view, two thousand years apart

Each Now panorama is an edit of the finished Then panorama with the same model, so the geometry and framing stay identical and the toggle reads as time, not as a different photograph. In the browser a single shader holds both layers, revealed through a wipe, a lens or a full ripple. For the 3D sites the Now layer is a second set of geometry in the same world coordinates.

The Now views are edits of the Then panoramas that follow the bibles' description of the sites today. They are not photographs.

Philae, around AD 100.
The same view on Agilkia today.

05 · The walkable sites

Two monuments built from numbers

The pyramid is generated by script from measured constants: 217 courses, the doorway on course 19 at 16.97 m, the Grand Gallery at 26.28 degrees and 8.6 m high with seven corbels, every passage in the plane x = 7.29 m. The interior is a union of void volumes at true scale, in the same world coordinates as the plateau. PATINA Material and Scenario Texture supplied the stone; the course atlases and block joints were drawn in code.

The hall has all 134 columns as lathe profiles with painted zones, architraves, a roof with light slots and a clerestory with nine-slot grilles. A Python ray-cast solver in Blender computed sun, sky and one bounce per vertex on the CPU and stored it as vertex colour, since no Cycles bake could run. Eight distinct wall relief panels follow the documented programme: Seti I raised relief in the north, Ramesses II sunk relief in the south.

Painted stop views

At every stop of the two walkable sites the live scene dissolves into a painted panorama made from the same camera. A 360 guide was rendered from the 3D scene at each stop, then repainted by GPT Image 2.5 Sunburst (it beat Scenario Skybox Flux in a shoot-out) with prompts naming what must stay: the outline, the number of people, the lamps. Topaz Image Upscale took each plate to 8192 px, in Then and Today versions. Walk, and the painting crossfades back into the 3D model.

Khufu, the white pyramid: rendered from the 3D scene.
The painted stop view.
Karnak, the clerestory: rendered from the 3D scene.
The painted stop view.
The Grand Gallery by lamplight: the painted stop view.

Figures came from GPT Image 2.5 Sunburst references and an image-to-3D shoot-out: Meshy 7.1 won the Karnak bearer and barque, Hitem3D 3.0 won the Khufu overseer and hauler, Meshy the Khufu priest. Phones load lighter scenes: Karnak drops from 871,848 to 234,793 triangles.

06 · Hieroglyphs

Real names, set in type

Image models write convincing pseudo-glyphs. Every readable name in the visits is therefore typeset from the Noto Sans Egyptian Hieroglyphs font (Unicode, SIL Open Font License), with each code point checked against the bibles and Unicode 16.0. Philae's research caught two wrong code points in the brief itself: the seat sign is U+132A8, not U+133E8.

Generated reliefs remain in the panoramas and walls, where they are kept at viewing distance and never described as readable. At Karnak the titulary strips and the cartouches on columns and architraves are real; the columns of text inside painted scenes are decorative.

The cartouches of Seti I and Ramesses II, typeset from Noto Sans Egyptian Hieroglyphs.

07 · Voice and score

Two shoot-outs for one guide

A first shoot-out (11 takes, 4 models, 7 voices) chose a steady narrator for pronunciation: ElevenLabs Multilingual 2, George. A second one, later the same day, replaced it with a more expressive delivery: 8 candidates and 21 takes of the same three passages from the real scripts (factual with dates, awe, intimate), each measured for prosody.

Google Lyria 3 Pro won a four-way test: the only take that was continuous, pulsed and loopable.

ElevenLabs Music Extend v2.5 continued a 12 s slice of the recorded hub motif in a new orchestration per site, then the slice was cut. The five beds share the theme acoustically; a text-only run wrote new melodies.

Opening film key frames: Giza, Karnak, the Valley of the Kings, Philae, Abu Simbel.

08 · Scenario models

Fourteen models, each chosen for its job

Every Scenario job was logged with model, purpose, CU, asset IDs and whether it was kept: 542 ledger entries in six files. No model was chosen as a default; each came out of a recommendation, a newest-first search and, where two were close, a shoot-out.

Driven by agents through Scenario MCP with the Scenario Skills. Connect Scenario MCP to your own agent and install the skills to work the same way.

09 · Cost

36,482 Scenario credits in the ledgers

Summed from the six project ledgers, including rejected takes, shoot-out losers and the painted stop views. Seedance 2.5 is the largest single item. No other paid service was used.

Great Pyramid7,531
Karnak4,953
Tutankhamun2,754
Philae2,760
Abu Simbel3,574
Voice, score, films14,910
Total36,482

10 · Honest limits

What this reconstruction is not

  • No laser scans. Geometry comes from published measurements and plans. Khufu's per-course heights are synthetic between anchored courses; Karnak's heights above 21 m are uncertain.
  • AI-painted reliefs. Wall paintings and inscriptions in the panoramas are generated and illustrative at viewing distance. Only the typeset cartouches are readable text.
  • Estimated positions. Philae's buildings are placed within about 10 m and 5 degrees, pending the Lyons 1896 plans.
  • No Cycles bake. The shared render queue never opened. Karnak carries a CPU light solution and Khufu realtime lighting; the painted plates carry the photographic light at each stop.
  • Choices, stated. Karnak's twenty bearers are shown unmasked (the hall's relief shows them masked); no one knows how many carried the boat. The narration says both.
  • Listening. Voice takes were measured for prosody and checked by whisper; music beds were judged by spectrogram and loudness.

Take the method with you

Visit the five sites, switch between Then and Today at every stop, and download the agent skill that recreates this workflow for any heritage site.

Typefaces: Cormorant Garamond and Source Sans 3 (SIL Open Font License). Hieroglyphs: Noto Sans Egyptian Hieroglyphs (SIL Open Font License). 3D engine: three.js (MIT). Sources are listed per site under Learn.

Learn

Learn

Sources

    How it was made

    Measured research first, then Scenario models for panoramas, textures, voices and music, driven through Scenario MCP.

      Make your own

      Privacy:

      Keyboard

      Space
      Play or pause the guided tour
      N P
      Next or previous stop
      T
      Switch between Then and Today
      L
      Compare then and today (360 visits)
      ← → ↑ ↓
      Look around (360) or walk and turn (3D)
      W A S D
      Walk (3D), hold Shift to stride
      + −
      Zoom (360)
      C
      Captions on or off
      M
      Sound on or off
      F
      Fullscreen
      Esc
      Close a panel
      Esc Esc
      Return to the map of all sites
      ?
      This list