The audience is the stage.

1 · The anchor. The AR set is raised over something real, and the anchor image is how the phone recognises it. It can be a drawing printed and handed to the audience, a photograph of an object placed in the venue, or a photograph of the stage itself, taken from where the audience will stand. Any image works, but give it detail and contrast: at least 640 pixels across, and visually distinct from every other set's anchor in the show — two similar anchors make the phones guess which AR set to raise. Measure the real thing now: the width the anchor has in the venue (a printed A4 sheet is 0.21 m — the default) is what sets the scale of everything.

2 · The model. Build the AR set in Blender (or any tool that exports glTF), in real metres, in Blender's own Z-up axes. Imagine the anchor lying flat at the origin: the set stands on and around it, facing −Y — towards the viewer. An AR set anchored to a sheet in the audience's hands is small by design; one anchored to the stage is built at stage size. Name every surface that will carry video (screen-left, backdrop…) and UV-unwrap those surfaces — video lands through texture coordinates, and a mesh without them shows nothing.

3 · The projections. For each video surface, tell the AR set which file goes there. The cleanest way is in the model itself: add a custom property named video to the surface's object in Blender, with the video's filename as its value. Or skip properties and simply name each video file after its surface (screen-left.mp4). Or say it in ar-set.json — see below.

4 · Export. glTF Binary (.glb), one file — a .gltf with separate textures is refused. Two export settings matter: untick “+Y Up” (the phones expect the file Z-up, exactly as Blender has it, and do the turn themselves), and tick Include ▸ Custom Properties if the video bindings live in the model.

5 · The folder. One folder, named after the AR set: the anchor image, the .glb, the videos, and optionally a ar-set.json. Drop it on the Dashboard's delivery table. The import sheet checks everything — the anchor's size, the model's surfaces, every binding — and says plainly what is missing or ambiguous; nothing is sent until the errors are gone. Then Send, and the AR set travels to the house like any other file. Start raises it; Stop strikes it.

⇩ starter-set.zip — a complete working AR set with a README
The demo show's AR set: anchor, metre-scale .glb with two named screens, two videos and a fully spelled-out ar-set.json. Take it apart and replace every piece with your own.

One folder holds one AR set. It needs the anchor image — a drawing to be printed and handed out, or a photograph of the object or stage the set will stand on (JPEG or PNG, at least 640 pixels across; detailed, high-contrast views track best) — and the set itself as a single .glb, authored in metres (a .gltf with separate textures will be refused). Videos beside them (.mp4, .mov, .m4v) are projected onto the model's surfaces: say which goes where in the model's own node properties (a custom property named “video” in Blender), in a ar-set.json beside it, or simply by naming the video after the surface. An optional ar-set.json can also name the anchor and model among other files and give two numbers: “widthMeters”, the real-world width of the anchor as the audience sees it (0.21 — an A4 sheet — if unsaid), and “scale”, a correction for models not authored in metres (1 if unsaid). Both can also be adjusted in the import sheet. Keep an AR set well under 60 MB — every phone pulls all of it during walk-in.

A small JSON file beside the model, for what the files alone cannot say. Every key is optional — a folder with one obvious drawing, one .glb and self-naming videos needs no ar-set.json at all. The folder's name becomes the AR set's id and title.

{
  "marker":      "three-sisters.jpg",
  "model":       "stage.glb",
  "widthMeters": 0.297,
  "scale":       1,
  "projections": {
    "screen-left":  "clouds-left.mp4",
    "screen-right": "clouds-right.mp4"
  }
}

marker — which image is the anchor, when the folder holds more than one. Without it, the first image alphabetically is used, with a warning.

model — which .glb is the AR set, when there are several.

widthMeters — the width the anchor really has in the venue, in metres: 0.21 (the default) for a printed A4 sheet, 0.297 for A3, or the measured width of the object or stage view photographed. The whole AR set scales from this one number — the same image printed at two sizes is two different anchors, so measure the real thing.

scale — a correction for models not authored in metres: the multiplier that brings one model unit to one metre. 1 (the default) for a correctly authored file.

projections — surface name → video filename, for models whose bindings are not in the file itself. Bindings written in the model's own custom properties win over ar-set.json; a video named after its surface is the fallback when neither says.