Deterministic GLB meshes with known geometry and triangle counts for regression-testing mesh import/export, decimation, repair, and validation pipelines.
High-resolution sphere mesh — the ground truth for scoring a decimation / simplification tool against the low-poly member of this group.
Low-poly simplification of the same sphere shape (140 triangles vs 1728). Compare against the dense ground truth to measure decimation error.
High-resolution torus mesh — the ground truth for scoring a decimation / simplification tool against the low-poly member of this group.
Low-poly simplification of the same torus shape (196 triangles vs 1600). Compare against the dense ground truth to measure decimation error.
High-resolution vase mesh — the ground truth for scoring a decimation / simplification tool against the low-poly member of this group.
Low-poly simplification of the same vase shape (100 triangles vs 480). Compare against the dense ground truth to measure decimation error.
High-resolution cylinder mesh — the ground truth for scoring a decimation / simplification tool against the low-poly member of this group.
Low-poly simplification of the same cylinder shape (36 triangles vs 192). Compare against the dense ground truth to measure decimation error.
Watertight box mesh — the intended result for a mesh-repair / hole-filling tool applied to the broken member of this group.
Deliberately broken box mesh (missing top face) — feed to a mesh-repair tool and compare the result against the watertight ground truth.
Watertight tube mesh — the intended result for a mesh-repair / hole-filling tool applied to the broken member of this group.
Deliberately broken tube mesh (open, uncapped ends) — feed to a mesh-repair tool and compare the result against the watertight ground truth.
Watertight sphere mesh — the intended result for a mesh-repair / hole-filling tool applied to the broken member of this group.
Deliberately broken sphere mesh (open hole at the base) — feed to a mesh-repair tool and compare the result against the watertight ground truth.
A tube skinned to a two-joint chain with inverse bind matrices, unsigned-byte JOINTS_0 and float WEIGHTS_0, plus a three-second rotation animation on the elbow. The shipped models category had no skinned mesh at all; this is the smallest complete one — rig, bind pose, weights and clip in a single self-contained GLB.
The same two-joint skinned bend as readable glTF JSON with its buffer embedded as a base64 data URI — the file to open in a text editor when a GLB's skin, sampler or inverse bind matrices are not doing what you expect.
A longer tube skinned to a four-joint chain, every joint but the root animated, so the whole limb curls rather than hinging once. Exercises joint-matrix accumulation down a hierarchy — the step that a two-bone rig cannot catch.
One cube on one node, with translation, rotation and scale each driven by a five-key sampler that is linearly interpolated. Three files differing only in `interpolation` — the cleanest way to prove an animation player implements all three modes rather than treating everything as linear.
One cube on one node, with translation, rotation and scale each driven by a five-key sampler that is held (STEP) — the value jumps at each key. Three files differing only in `interpolation` — the cleanest way to prove an animation player implements all three modes rather than treating everything as linear.
One cube on one node, with translation, rotation and scale each driven by a five-key sampler that is cubic-spline, with in/out tangents stored around each value. Three files differing only in `interpolation` — the cleanest way to prove an animation player implements all three modes rather than treating everything as linear.
A single animation whose four channels drive two nodes at once: the parent's translation, rotation and scale, and a spinning child that inherits all of it. The fixture for checking that channels are applied per node and per path rather than collapsed into one track.
A 16x16 plane with two POSITION morph targets — a sine ripple and a centre bulge — both at weight 0 in the bind pose. Drive the weights yourself to check that a loader adds the deltas to the base positions instead of replacing them.
The same two-target morph plane as readable glTF JSON, so the `targets` array, the mesh-level `weights` and the target accessors can be inspected without a hex editor.
The two-target plane with an animation channel whose target path is `weights` — five keyframes, ten scalars, because a weights sampler stores one value per target per key. The case that breaks players which assume every channel output is a vec3 or a quaternion.
Morph targets that displace both positions and normals, with the first target already at weight 0.3 in the bind pose. Shading that stays flat while the surface ripples is the symptom of a loader that morphs positions and forgets the normal deltas.
Eight POSITION morph targets on one primitive — bumps arranged around a circle, so every active weight is visible independently. Eight is where many real-time renderers stop supporting extra targets; this is the fixture that says whether yours does.
A morph target stored as a sparse accessor: only 25 of 289 vertices carry a delta, and the accessor has no bufferView at all, so every other vertex reads as zero. Sparse accessors are the least-implemented corner of the glTF spec and the usual cause of a target that appears to do nothing.
The same sphere with its buffer inlined as a base64 data URI: one self-contained text file, at roughly a third more bytes than the binary it encodes. The middle point of the three storage strategies in this group.
The third storage strategy for the identical sphere: a GLB whose buffer is the BIN chunk, with no URI at all. Compare the three file sizes in this group to see exactly what base64 and JSON whitespace cost.
Position, normal and UV packed into ONE bufferView with byteStride 32 and three accessors at offsets 0, 12 and 24 — the GPU-friendly layout, and the one a naive parser mis-reads because it assumes each accessor owns its view. Twinned here with a tightly-packed version of the same sphere.
The same sphere with every attribute in its own bufferView and no byteStride — the layout most exporters emit and most parsers assume. The control against which the interleaved twin in this group is diffed.
An unquantised sphere: 32-bit floats for both positions and normals, 24 bytes of attribute data per vertex. The uncompressed reference for the quantized and gzip twins in this group.
The same sphere with positions as normalised 16-bit and normals as normalised 8-bit integers under KHR_mesh_quantization — 66.6% of the float32 twin's bytes. The extension is in extensionsRequired, so a loader without support must refuse the file rather than render garbage.
A welded, manifold, consistently-wound unit cube: 8 vertices, 12 triangles, 18 edges each shared by exactly two faces, and a signed volume of exactly 1.0. This is the answer key — each defect fixture in the degenerate/ subcategory is this mesh with exactly one thing wrong.
The clean reference cube as a GLB, so the same baseline is available to glTF tooling and to the site's 3D preview. Identical topology to the OBJ twin in this group.
The reference cube plus four triangles with no area, one for each way a mesh can have one: three collinear points, two coincident vertices at different indices, a repeated index, and all three indices identical. Their face normals are undefined, which is where downstream tools produce NaN.
The same four degenerate triangles inside a structurally valid GLB. The file passes every glTF schema check — nothing in the format forbids a zero-area face — so this is the fixture for geometry validation as opposed to document validation.
An ASCII STL of the reference cube with two extra facets that have no area and declare a null (0 0 0) normal — exactly what CAD tessellators emit at a sliver. Slicers either drop them, warn, or divide by zero; this is how you find out which.
The reference cube with all eight corners stored twice at bit-identical coordinates: the first six triangles use one copy, the last six use the other. It renders as a closed cube but is topologically two disconnected surface patches with a split seam, which is why smoothing, subdivision and boolean operations all misbehave on it.
The duplicated-corner cube as ASCII PLY — the format scanners and photogrammetry tools emit, and the one where unwelded duplicates are most common because each captured point arrives independently.
The duplicated-corner cube as a GLB. Because glTF stores every attribute per vertex, the duplication doubles the position and normal buffers for a shape that needs eight corners — the cheapest optimisation an asset pipeline can make, and the one this fixture measures.
The reference cube with a two-triangle fin welded along one top edge, so that edge is shared by three faces instead of two. Non-manifold edges have no defined surface orientation, which is why slicers refuse them and why subdivision and offsetting produce garbage.
The finned cube as a GLB. It renders without complaint in any viewer — rasterisation never asks about adjacency — so this is the fixture that separates tools which actually analyse topology from tools which merely draw triangles.
The reference cube with three of its six cube faces (six of its twelve triangles) wound clockwise instead of counter-clockwise, and `vn` records that follow the reversed winding into the solid. The classic symptom is a model that looks half-invisible with backface culling on and half-black with it off.
The mixed-winding cube as a GLB with a default single-sided material, so the reversed faces are culled and you can see straight through the model. Turning on doubleSided hides the defect without fixing it, which is the trap this fixture is for.
The same unit sphere at 16,128 triangles — level 0 of a four-step ladder that spans 202x from top to bottom. Because every level is generated analytically from the same surface rather than decimated, the ladder is an exact reference: any deviation you measure is your tool's, not the fixture's.
The same unit sphere at 3,968 triangles — level 1 of a four-step ladder that spans 202x from top to bottom. Because every level is generated analytically from the same surface rather than decimated, the ladder is an exact reference: any deviation you measure is your tool's, not the fixture's.
The same unit sphere at 960 triangles — level 2 of a four-step ladder that spans 202x from top to bottom. Because every level is generated analytically from the same surface rather than decimated, the ladder is an exact reference: any deviation you measure is your tool's, not the fixture's.
The same unit sphere at 80 triangles — level 3 of a four-step ladder that spans 202x from top to bottom. Because every level is generated analytically from the same surface rather than decimated, the ladder is an exact reference: any deviation you measure is your tool's, not the fixture's.
The densest level of the ladder as a binary little-endian PLY — 8,320 vertices and 16,128 faces, for the mesh-processing tools (MeshLab, CloudCompare, Open3D) that speak PLY rather than glTF.
The coarsest level as plain-text OBJ — 80 triangles you can read end to end. Small enough to paste into a bug report, which is what makes it useful when a pipeline mangles winding or normals and you need the smallest reproducer.
One box mesh referenced by a 24-deep chain of nodes, each adding a translation, an 11-degree rotation and a 0.94 scale, so the result spirals. Any error in matrix accumulation compounds visibly by the tip instead of hiding in rounding.
The 24-level spiral as readable JSON, so the `children` arrays and per-level TRS can be traced by eye. Useful when a converter flattens or re-parents a hierarchy and you need to see exactly where.
Six cubes in a ring whose nodes carry a column-major 4x4 `matrix` and no translation/rotation/scale keys — the encoding a DCC exporter uses when a transform cannot be decomposed. Paired with a TRS twin of the identical ring.
The same six-cube ring with translation, quaternion rotation and scale spelled out per node. Read it next to the matrix twin to check a decomposition routine, or to confirm that your quaternion component order really is xyzw.
An 8x8 field of columns built from a single 12-triangle mesh referenced by 64 nodes with seeded random heights. Core-glTF instancing needs no extension: the file is tiny, and the interesting question is whether your renderer batches the 64 draws or issues them one by one.
Twelve nodes reference one sphere mesh, in three groups whose parents each apply a different rotation and scale. The fixture for importers that copy geometry per node: a correct import keeps one mesh resource with twelve transforms, a naive one produces twelve meshes.
Thirty-two copies of one cube expressed as instance attribute accessors on a single node under EXT_mesh_gpu_instancing — the extension that moves instancing off the scene graph and onto the GPU. A viewer without support shows exactly one cube at the origin, which is the tell.
Three named scenes over three meshes, with the asset's `scene` property pointing at index 1 rather than 0. Loaders that hardcode scene 0 show the wrong content and never report an error — this is the file that catches them.
The three-scene asset as readable JSON, where the top-level `"scene": 1` is one line you can point at in a bug report. Also the fixture for round-tripping multi-scene assets through tools that keep only one scene.
A 1.12 m signpost with its coordinates written in metres — so this file's numbers read up to 1.12. glTF has no unit field: the spec simply says metres, which means a wrongly-scaled export is indistinguishable from a genuinely enormous model until something is measured.
A 1.12 m signpost with its coordinates written in centimetres — so this file's numbers read up to 112. glTF has no unit field: the spec simply says metres, which means a wrongly-scaled export is indistinguishable from a genuinely enormous model until something is measured.
A 1.12 m signpost with its coordinates written in millimetres — so this file's numbers read up to 1120. glTF has no unit field: the spec simply says metres, which means a wrongly-scaled export is indistinguishable from a genuinely enormous model until something is measured.
The same signpost as OBJ with coordinates in metres. OBJ has no unit record of any kind, so the two OBJ files in this group are byte-different and semantically identical — the ambiguity that every OBJ import dialog's scale box exists to resolve.
The same signpost as OBJ with coordinates in millimetres. OBJ has no unit record of any kind, so the two OBJ files in this group are byte-different and semantically identical — the ambiguity that every OBJ import dialog's scale box exists to resolve.
The signpost in millimetres as ASCII STL. STL has no unit field either, but every slicer assumes millimetres — which is why an STL exported in metres arrives as a 1 mm speck and one exported in inches arrives 25.4x too big.
The signpost authored +Y up, which is what the glTF specification mandates and what every conformant viewer assumes. The conformant half of the axis pair: load it, confirm the post stands, and you have calibrated your viewer before testing the Z-up twin.
The same signpost as COLLADA with `<up_axis>Y_UP</up_axis>` and `<unit meter="1" name="metre"/>` in its asset block. Unlike glTF, OBJ and STL, COLLADA states its convention — so an importer that still gets the orientation wrong has no excuse.
The signpost as ASCII PLY with +Y up. PLY has no axis or unit element, so the convention survives only as a header comment that no parser reads — the reason scan data so often arrives rotated.
The signpost as USDZ with `upAxis = "Y"` and `metersPerUnit = 1` in the stage metadata. USD is the only format in this group that declares BOTH the axis and the scale, which is exactly why it is the one AR pipelines standardised on.
The same signpost authored +Z up — what a Blender or 3ds Max glTF export produces when the axis conversion is skipped. glTF has no field in which to declare this, so the file is indistinguishable from a model that genuinely lies on its back, and every conformant viewer will show it that way.
The same signpost as COLLADA with `<up_axis>Z_UP</up_axis>` and `<unit meter="1" name="metre"/>` in its asset block. Unlike glTF, OBJ and STL, COLLADA states its convention — so an importer that still gets the orientation wrong has no excuse.
The signpost as ASCII PLY with +Z up. PLY has no axis or unit element, so the convention survives only as a header comment that no parser reads — the reason scan data so often arrives rotated.
The signpost as USDZ with `upAxis = "Z"` and `metersPerUnit = 1` in the stage metadata. USD is the only format in this group that declares BOTH the axis and the scale, which is exactly why it is the one AR pipelines standardised on.
Real deterministic FBX tetrahedron source with four vertices, four triangular faces, and unit bounds for Convert v2 import coverage. Stable P8 artifact p8-convert-model-fbx.
JSON semantic reference for the FBX tetrahedron, publishing vertex order, face indices, bounds, and the closed-mesh contract. Stable P8 artifact p8-convert-model-fbx-reference.
Real deterministic 3DS tetrahedron source with four vertices, four triangular faces, and unit bounds for Convert v2 import coverage. Stable P8 artifact p8-convert-model-3ds.
JSON semantic reference for the 3DS tetrahedron, publishing vertex order, face indices, bounds, and the closed-mesh contract. Stable P8 artifact p8-convert-model-3ds-reference.
Real deterministic AMF tetrahedron source with four vertices, four triangular faces, and unit bounds for Convert v2 import coverage. Stable P8 artifact p8-convert-model-amf.
JSON semantic reference for the AMF tetrahedron, publishing vertex order, face indices, bounds, and the closed-mesh contract. Stable P8 artifact p8-convert-model-amf-reference.
Real deterministic LWO tetrahedron source with four vertices, four triangular faces, and unit bounds for Convert v2 import coverage. Stable P8 artifact p8-convert-model-lwo.
JSON semantic reference for the LWO tetrahedron, publishing vertex order, face indices, bounds, and the closed-mesh contract. Stable P8 artifact p8-convert-model-lwo-reference.
Real deterministic VTK tetrahedron source with four vertices, four triangular faces, and unit bounds for Convert v2 import coverage. Stable P8 artifact p8-convert-model-vtk.
JSON semantic reference for the VTK tetrahedron, publishing vertex order, face indices, bounds, and the closed-mesh contract. Stable P8 artifact p8-convert-model-vtk-reference.
Real deterministic VOX source with four occupied voxel coordinates, palette indices, and unit occupied-coordinate bounds for Convert v2 import coverage. Stable P8 artifact p8-convert-model-vox.
JSON semantic reference for the VOX fixture, publishing occupied coordinates, palette indices, empty face topology, bounds, and the voxel contract. Stable P8 artifact p8-convert-model-vox-reference.
Real deterministic DRC tetrahedron source with four vertices, four triangular faces, and unit bounds for Convert v2 import coverage. Stable P8 artifact p8-convert-model-drc.
JSON semantic reference for the DRC tetrahedron, publishing vertex order, face indices, bounds, and the closed-mesh contract. Stable P8 artifact p8-convert-model-drc-reference.
We use Google Analytics and show ads via Adsterra. Non-essential cookies and ad scripts run only after you allow the matching categories. See our cookie policy.
