XGF (xeokit Geometry Format) Loader / Exporter

XGFLoader reads xeokit's native binary geometry format (.xgf) into a SceneModel, and XGFExporter writes a SceneModel back out as XGF.

This document describes the binary container, the two on-disk schema versions, and the pipelines a file follows from bytes to a live scene (and back).


XGF is xeokit's compact binary format for geometry, materials, and textures — the visual half of a model. It's positional, length- prefixed, and designed to be loaded in one fetch + a handful of typed-array views.

Two roles:

  • Visual half — the SceneModel — lives in .xgf.
  • Semantic half — the DataModel — typically lives in a separate .json sidecar (one of the xeokit datamodel / metamodel formats). Pair them by id and load them together.

The XGF loader's DataModel output is intentionally minimal: every SceneObject id also becomes a BasicEntity DataObject so the SceneModel ↔ DataModel pair is always queryable. Richer semantics come from a paired DataModel JSON.

  • Positions are quantised to 16-bit against per-geometry AABBs — no precision loss for typical model scales (mm precision at world sizes up to ~65 m per geometry; rescale your AABBs for bigger scenes).
  • Normals are oct-encoded into two uint16s per vertex, decoded in the vertex shader.
  • UVs are float32 so tiling values outside [0, 1] survive intact.
  • Edge indices ride alongside triangle indices so the renderer's wireframe overlay is free.
  • The format intentionally has no compression layer of its own — use HTTP transport-level compression (gzip / brotli) at delivery time.

Every XGF file (v1 or v2) is the same outer container — a version tag, an (offset, length) table, and a sequence of typed-array payloads laid out at those offsets:

   ┌──────────────────────────────────────────────────────────────┐
0..3 uint32 version (1, 2 or 3) │
├──────────────────────────────────────────────────────────────┤
4..7 uint32 entry[0].byteOffset
8..11 uint32 entry[0].byteLength
12..15 uint32 entry[1].byteOffset
16..19 uint32 entry[1].byteLength
│ ... (one (offset, length) pair per entry) │
├──────────────────────────────────────────────────────────────┤
entry[0] typed-array payload (padded to its BPE align) │
entry[1] typed-array payload
│ ... │
└──────────────────────────────────────────────────────────────┘
  • Little-endian. The reader byte-swaps on big-endian hosts (a one-time Uint16Array[0] === 1 probe decides whether swapping is needed).
  • Position alignment. Each payload starts at the smallest offset ≥ the running cursor that's an integer multiple of the payload's BYTES_PER_ELEMENT. Padding bytes between payloads are unused.
  • Positional layout. The reader and writer walk the same fixed list of entries in the same order. There's no name dictionary — adding or removing an entry is a schema-version-breaking change.
  • JSON-as-bytes. A handful of entries (e.g. eachObjectId, eachMaterialId) are JSON arrays of strings; they ride in the container as Uint8Array blobs the reader JSON.parses on demand.

XGF has shipped three schema revisions. The loader handles all of them; the exporter writes any via its version parameter.

Version Adds
v1 (1.0.0) base — geometry (positions, colors, indices, edge indices), per-geometry AABBs, modelling matrices, meshes referencing per-mesh inline RGBA colours, objects referencing meshes.
v2 (1.1.0, "XKT2") adds per-geometry normals + UVs; textures (image bytes + sampler params); materials with full PBR + alpha mode / cutoff; per-mesh material references (with inline RGBA as the fallback when no material is set).
v3 (1.2.0) adds 3D Gaussian Splatting geometry — per-splat scales (float xyz) and rotation quaternions (xyzw, byte-quantised, the .splat convention), plus the GaussianSplatsPrimitive type (5). Also restores per-vertex colorsCompressed on decode, which splats need for baked colour.

The older readers stay in the codebase so existing .xgf files keep loading.

The default exporter version is 1.0.0. Pass version: "1.1.0" to emit v2 (PBR / textures), or version: "1.2.0" to emit v3 — required to preserve Gaussian-splat geometry, since v1/v2 have no splat buffers.


Field order = pack order = read order. See versions/v2/XGFData_v2.ts for the authoritative TypeScript shape.

Entry Type Meaning
positions Uint16Array Quantised positions (R G B 16-bit per vertex). Decode against the geometry's AABB: p = uint * (max - min) / 65535 + min.
colors Uint8Array Per-vertex RGBA (4 bytes/vertex).
indices Uint32Array Triangle / line indices.
edgeIndices Uint32Array Wireframe edge indices.
aabbs Float32Array Per-geometry AABBs (six floats: minX minY minZ maxX maxY maxZ). Multiple geometries may share an AABB by pointing at the same base.
normals Uint16Array v2 only. Octahedral RG16UI normals, two values per vertex. Geometries without normals occupy zero range.
uvs Float32Array v2 only. RG32F UVs, two values per vertex. Floats (not quantised) so tiling values round-trip.

eachGeometry* is a per-geometry array; each entry is the base index into the global blob it points to. The geometry's slice runs from its base to (the next geometry's base) or (the global blob's length).

Entry Type Meaning
eachGeometryPositionsBase Uint32Array Base into positions.
eachGeometryColorsBase Uint32Array Base into colors.
eachGeometryIndicesBase Uint32Array Base into indices.
eachGeometryEdgeIndicesBase Uint32Array Base into edgeIndices.
eachGeometryNormalsBase Uint32Array Base into normals, or 0xffffffff if this geometry has no normals.
eachGeometryUVsBase Uint32Array Base into uvs, or 0xffffffff if no UVs.
eachGeometryPrimitiveType Uint8Array 0 Triangles, 1 Solid, 2 Surface, 3 Lines, 4 Points.
eachGeometryAABBBase Uint32Array Base into aabbs (multiple of 6).
Entry Type Meaning
matrices Float64Array Concatenated 4×4 column-major modelling matrices. Each mesh points at one.
Entry Type Meaning
textureData Uint8Array Concatenated encoded image bytes (PNG / JPEG / GIF or opaque transcoded buffer).
eachTextureDataBase Uint32Array Base into textureData.
eachTextureMediaType Uint8Array 0 PNG, 1 JPEG, 2 GIF, 255 opaque transcoded buffer (treat as raw bytes; not decoded at load time).
eachTextureWidth Uint16Array Per-texture pixel width.
eachTextureHeight Uint16Array Per-texture pixel height.
eachTextureSampler Uint8Array Five bytes per texture: minFilter, magFilter, wrapS, wrapT, wrapR. Each is a small-integer code (1–9) matching the xeokit/base/constants sampler enums.
eachTextureId string[] Per-texture string ID (JSON blob).
Entry Type Meaning
eachMaterialPBR Uint8Array Eight bytes per material: R, G, B, opacity, roughness, metallic, alphaMode, alphaCutoff. All values quantised to [0, 255] from [0, 1]. alphaMode is 0 OPAQUE, 1 MASK, 2 BLEND.
eachMaterialTextures Int32Array Five entries per material: colorTextureIndex, metallicRoughnessTextureIndex, normalsTextureIndex, occlusionTextureIndex, emissiveTextureIndex. Each is a 0-based index into eachTextureId, or -1 for "no texture".
eachMaterialId string[] Per-material string ID (JSON blob).
Entry Type Meaning
eachMeshGeometriesBase Uint32Array Index into the geometry list (each mesh references exactly one geometry).
eachMeshMatricesBase Uint32Array Index into matrices (multiple of 16).
eachMeshMaterialAttributes Uint8Array Inline RGBA fallback (4 bytes/mesh: R, G, B, opacity). Used when eachMeshMaterial is -1.
eachMeshMaterial Int32Array Index into the material array, or -1 to fall back to eachMeshMaterialAttributes. v1 has no materials; always falls back.
Entry Type Meaning
eachObjectId string[] Per-object string ID (JSON blob).
eachObjectMeshesBase Uint32Array Base into the per-mesh arrays. The object's meshes run from its base to (the next object's base) or (mesh count).

v1 is the same shape as v2 minus normals, uvs, eachGeometryNormalsBase, eachGeometryUVsBase, eachGeometryPrimitiveType, all texture entries, and all material entries. The v1 reader treats every mesh as inline-RGBA and emits TrianglesPrimitive geometries by default. See versions/v1/XGFData_v1.ts.


   .xgf bytes (ArrayBuffer)


XGFLoader.getVersionread uint32 [0]: "1" or "2"


versions/v{N}/parse.ts

├─ unpackXGFwalk the (offset, length) table,
byte-swap on big-endian hosts,
create typed-array views in place
│ (zero copies)


XGFData_v{N} payload


xgfToModelwalks objectsmeshesgeometries

├─ decode textures (v2)
PNG/JPEG/GIFcreateImageBitmapsceneModel.createTexture
opaquepass raw bytes through with `compressed: true`
emptyregister a 1×1 white pixel placeholder

├─ build materials (v2)
PBR + alpha + texture referencessceneModel.createMaterial

├─ for each object:
for each mesh:
ensure geometry created oncesceneModel.createGeometryCompressed
│ (per-vertex slices come from `eachGeometry*Base` arrays;
normals/UVs slices end at the *next* geometry with
non-sentinel base, since geometries without
normals/UVs are sparsely indexed)
resolve mesh transform from matrices[base..base+16]
resolve material (v2) or inline RGBA (v1 / unset)
sceneModel.createMesh
sceneModel.createObject({id, meshIds, layerId})
dataModel.createObject({id, type: "BasicEntity"})
dataModel.createRelationship({type: "BasicAggregation", …})


SceneModel + DataModel populated

Every async-loop site calls yieldToHost(signal) at a coarse cadence (every 4 textures, every 32 objects). Two effects:

  • Progressoptions.onProgress fires with {phase, current, total} at the same cadence.
  • AbortSignalyieldToHost checks signal.aborted and throws DOMException("Aborted", "AbortError") on cancel. Worst-case abort latency is one yield interval (≈16 ms).

PNG/JPEG/GIF blobs are decoded through createImageBitmap so the GPU atlas can sample them directly. Outside a browser (Node test runs, worker contexts without createImageBitmap) the loader will throw — the path assumes a browser. For pre-decoded ("opaque transcoded") textures the bytes pass through unchanged as a SceneTexture buffer; the runtime / transcoder pipeline handles upload.


   SceneModel


modelToXGF (v1 or v2)

├─ flatten geometriesconcatenated positions / colors /
indices / edge indices, build base arrays
├─ flatten matricesone per mesh
├─ encode textures (v2)
imageBitmapcanvas.toBlob("image/png") → bytes
bufferspassed through unchanged
├─ build per-material PBR bytes + texture-index references (v2)
├─ build per-mesh material index OR inline RGBA fallback
└─ build per-object id + mesh-base arrays


XGFData_v{N} payload


packXGFpositional pack; byte-swap on big-endian,
pad each payload to its BPE alignment,
embed JSON arrays as Uint8 blobs

.xgf bytes (ArrayBuffer)

The pack ↔ unpack halves walk the same positional list in the same order — adding, removing, or reordering an entry is a schema-version-breaking change that requires bumping the version tag in XGF_INFO.xgfVersion and a matching reader.

Geometry data round-trips through createGeometryCompressed (load side) and the corresponding v2 packer (write side) without re-decompressing. Per-vertex positions, normals, UVs and indices are already quantised in the SceneModel, so the writer mostly concatenates buffers and emits base arrays.

If options.coordinateSystem is passed to the exporter, getMeshWorldMatrix bakes the SceneModel-space → target-space transform into each mesh's modelling matrix before pack. The XGF container itself doesn't carry a coordinate-system field — the caller specifies it at load time via the SceneModel.

imageBitmap-backed SceneTextures are re-encoded to PNG bytes via an OffscreenCanvas (or a <canvas> fallback). buffers-backed SceneTextures pass through verbatim, with mediaType: 255 flagging the opaque-transcoded case.


   uint16 = round((double - min) * 65535 / (max - min))
double = uint16 * (max - min) / 65535 + min

Applied per axis, per geometry. The AABB ships in aabbs and is referenced by eachGeometryAABBBase. Multiple geometries can share an AABB to save space when they cover the same volume.

Practical implication: keep each geometry's AABB tight to its actual extent. At a 1-metre AABB you get ~15 µm precision; at a 100-metre AABB you get ~1.5 mm.

Each normal is folded to the unit-octahedron's 2D parameterisation, then quantised to two 16-bit unsigned ints. Decoded in the vertex shader — the renderer's unoctEncodedVec3 does the inverse. Cost: 2 bytes/normal vs 12 bytes for raw float3.

Floats, not quantised. Tiling values outside [0, 1] (e.g. for repeat-mapping a brick texture across a wall) need to survive intact, which they wouldn't through a 0..1-clamped quantisation.


import {Scene} from "@xeokit/sdk/model/scene";
import {Data} from "@xeokit/sdk/model/data";
import {XGFLoader} from "@xeokit/sdk/formats/xgf";

const scene = new Scene();
const data = new Data();

const sceneModel = scene.createModel({id: "myModel"}).value!;
const dataModel = data.createModel({id: "myModel"}).value!;

const fileData = await (await fetch("./model.xgf")).arrayBuffer();

await new XGFLoader().load({
fileData,
sceneModel,
dataModel,
});

The loader's fileDataType is "arraybuffer". Pass a DataModel to also receive the loader's minimal BasicEntity graph keyed by SceneObject id; omit it if you'll pair the SceneModel with a hand-authored DataModel from a JSON sidecar.

import {XGFExporter} from "@xeokit/sdk/formats/xgf";

const arrayBuffer = await new XGFExporter().write({
sceneModel,
version: "1.1.0", // omit for default "1.0.0"
});

The exporter's fileDataType is "arraybuffer". The dataModel parameter is accepted by the base ModelExporter API but XGF is a geometry-only format — semantic data should be written separately through the datamodel JSON exporter.

The canonical streamed-model payload is .xgf + .json loaded together:

import {DataModelImporter} from "@xeokit/sdk/formats/datamodel";
import {XGFLoader} from "@xeokit/sdk/formats/xgf";

await Promise.all([
new DataModelImporter().load({
fileData: await (await fetch("./model.json")).json(),
dataModel,
}),
new XGFLoader().load({
fileData: await (await fetch("./model.xgf")).arrayBuffer(),
sceneModel,
dataModel, // safe to share — the XGF BasicEntity graph merges
// with the JSON-authored DataObjects by id
}),
]);

  • Animation / skinning. XGF is a static-geometry format. Time- varying transforms aren't representable.
  • Cameras, lights. The format describes meshes + materials, not the surrounding scene.
  • NURBS / B-rep / parametric primitives. Geometry must be tessellated to triangles (or lines / points) before export.
  • Section planes. A runtime view-side concept; not in the file.
  • Semantic structure. DataObjects / Relationships / PropertySets live in the paired DataModel JSON; XGF only carries object IDs so the two halves can pair up.
  • Coordinate system metadata. The runtime caller specifies the coordinate system via SceneModel.coordinateSystem; the file itself is coordinate-system-agnostic.
  • Pre-encoded compression layer. No built-in zstd / draco / meshopt step. Use HTTP-layer compression (gzip / brotli) for wire efficiency.

formats/xgf/
├── README.md (this file)
├── XGFLoader.ts ModelLoader subclassversion dispatch
├── XGFExporter.ts ModelExporter subclassversion dispatch
├── index.ts module re-exports
└── versions/
├── v1/
│ ├── XGF_INFO.ts {xgfVersion: 1}
│ ├── XGFData_v1.ts v1 payload type
│ ├── parse.ts load pipeline entryunpackxgfToModel
│ ├── encode.ts write pipeline entrymodelToXGFpack
│ ├── unpackXGF.ts ArrayBufferXGFData_v1 (positional, BE-aware)
│ ├── packXGF.ts XGFData_v1ArrayBuffer
│ ├── xgfToModel.ts XGFData_v1SceneModel + DataModel
│ └── modelToXGF.ts SceneModelXGFData_v1
└── v2/
├── XGF_INFO.ts {xgfVersion: 2}
├── XGFData_v2.ts v2 payload type (adds normals, UVs, textures, materials)
├── parse.ts load pipeline entry
├── encode.ts write pipeline entry
├── unpackXGF.ts ArrayBufferXGFData_v2
├── packXGF.ts XGFData_v2ArrayBuffer
├── xgfToModel.ts XGFData_v2SceneModel + DataModel
└── modelToXGF.ts SceneModelXGFData_v2