GLB Web Export

Take a .glb/.gltf that came out of Blender, a CAD tool, a photogrammetry pipeline, or a Pascal export, and turn it into an asset a browser can download and render quickly — without silently breaking the things the app depends on.

The deliverable is not a smaller file. It is a smaller file plus a report that proves nothing regressed.

When to Use

• A .glb/.gltf is too big, slow to load, or blows up GPU memory
• Preparing an asset for Three.js, R3F/drei, Babylon.js, or <model-viewer>
• Choosing between Draco, meshopt, and no geometry compression
• Choosing between KTX2/Basis, WebP, and resized PNG/JPEG for textures
• Auditing an asset before shipping it, or diagnosing why it loads slowly

When NOT to Use

• Editing or modelling geometry, or laying out/authoring a scene
• Generating a 3D model from text, an image, or a floor plan
• Converting from FBX/OBJ/STEP (do the conversion first, then use this skill)
• Runtime frame-rate problems unrelated to asset size (shaders, physics, effects)

Workflow

1. Measure  →  2. Decide  →  3. Apply  →  4. Verify  →  5. Report

Never skip step 1. An optimization with no measured before is a guess.

1. Measure

Keep the original. Every command below writes a new file; never optimize in place.

cp model.glb model.original.glb

Check the tool versions rather than assuming (package names and versions move):

npm view @gltf-transform/cli version
npm view gltfpack version

Inspect, then validate against the spec:

npx @gltf-transform/cli inspect model.glb
npx @gltf-transform/cli validate model.glb

inspect gives resolution and gpuSize per texture, per-mesh vertex/triangle counts, and the extension list. validate runs the Khronos validator (the gltf-validator npm package is a library with no CLI — npx gltf-validator fails with "could not determine executable to run"; gltf-transform validate wraps it, verified with @gltf-transform/cli@4.5.0 bundling gltf-validator@2.0.0-dev.3.10).

Then capture the numbers you will be judged on:

bash <skill-path>/scripts/glb-audit.sh model.glb

It reports raw/gzip/brotli bytes, texture VRAM, draw calls, triangles, world-space bounds in metres, named nodes, and required extensions. Pass a second path after optimizing to get a delta plus an automatic regression list.

Read inspect to answer: is this texture-heavy (large size/gpuSize), geometry-heavy (high vertex counts, f32 attributes), draw-call-heavy (many mesh primitives), or just dirty (unused materials, duplicate textures)? The answer decides step 2.

2. Decide

Geometry compression — pick exactly one

How to check the loader before choosing — do not guess:

• Three.js — grep the app for setMeshoptDecoder / setDRACOLoader. Absent means the extension will fail to load.
• R3F/drei — useGLTF enables Draco and meshopt by default; KTX2 needs extendLoader.
• <model-viewer> — Draco and KTX2 decoders load from a CDN by default; meshopt is not enabled by default.
• Babylon.js — Draco, meshopt, and KTX2 are built in and fetch decoders from the Babylon CDN.

Exact API names, defaults and doc URLs: references/loader-support.md.

Textures — usually the biggest win

Resize first. Halving each dimension quarters both bytes and VRAM, and no codec recovers a 4× texel reduction.

KTX2 wins on VRAM by roughly 6–8× over WebP at the same resolution, because PNG/JPEG/WebP/AVIF are fully decompressed in GPU memory while ETC1S/UASTC are not. On the fixture: 5.33 MiB VRAM (WebP) vs 0.67 MiB (ETC1S), same 1024px source.

Structural passes

Cheap, lossless, run these first: dedup (duplicate accessors/textures/ materials/meshes), prune (unreferenced properties), weld (merge bitwise identical vertices, producing an index buffer), quantize (f32 → i8/i16).

Structure-changing, opt in deliberately: instance (shared meshes → EXT_mesh_gpu_instancing, cuts draw calls, erases the per-node names), flatten (collapse the hierarchy), join (merge primitives — can increase size when meshes are reused; verified on the fixture: 12,882,440 → 12,890,228 B).

3. Apply

Two pipelines. Start with the first.

Name-preserving (default)

Keeps the scene graph and every node name an app might look up.

npx @gltf-transform/cli dedup  model.glb  s1.glb
npx @gltf-transform/cli prune  s1.glb     s2.glb          # add --keep-leaves to keep empty helper nodes
npx @gltf-transform/cli weld   s2.glb     s3.glb
npx @gltf-transform/cli resize s3.glb     s4.glb --width 1024 --height 1024
npx @gltf-transform/cli etc1s  s4.glb     s5.glb --quality 128   # or: webp s4.glb s5.glb --quality 85
npx @gltf-transform/cli meshopt s5.glb    model.web.glb --level high

Draw-call reduction (opt in)

Add before the compression step, only if the app does not resolve nodes by name:

npx @gltf-transform/cli instance s3.glb s3b.glb --min 5

One-liner

optimize bundles the above but its defaults are aggressive — it runs simplify, palette, flatten and join. On the fixture it dropped 15 of 16 named nodes and changed the triangle count (7,056 → 6,740). Use it only when the asset is a static prop with no app-visible structure:

npx @gltf-transform/cli optimize model.glb model.web.glb \
  --compress meshopt --texture-compress ktx2 --texture-size 1024

gltfpack is a good alternative and does geometry + textures in one pass. Use -kn -km to keep named nodes and materials. Flags: references/commands.md.

gltfpack -i model.glb -o model.web.glb -cc -tc -tl 1024 -kn -km -v

4. Verify

Re-run the audit against the original and check the regression list is empty:

bash <skill-path>/scripts/glb-audit.sh model.original.glb model.web.glb
npx @gltf-transform/cli validate model.web.glb

Then confirm each item in references/correctness-checklist.md. The non-negotiables:

• Units and scale. glTF units are metres and +Y is up (spec §Coordinate System and Units, fetched 2026-09-11). The audit's worldBounds.sizeMeters must be unchanged. Do not read inspect's bboxMin/bboxMax for this — after quantize/meshopt/draco it reports accessor-local bounds, and the fixture's harmless output shows -1,-1,-1 → 1,1,1 while the real world size is still 4.4 × 1 × 3.2 m.
• Node names and hierarchy. namedNodesLost must be empty, or every entry must be one the app provably does not use.
• Animations and skins. Counts must not drop. join/flatten will not move animated nodes, but prune removes animations that target nothing in a scene.
• Materials. Same count of distinct appearances, same alphaMode per material, KHR_materials_* extensions still listed, vertex colors (COLOR_0) still present.
• Loader can decode it. Every entry in extensionsRequired must be supported by the target loader, with its decoder wired up. Load the output in the real app before shipping.

5. Report

Report the measurement, not the effort. Template:

model.glb → model.web.glb

| Metric               | Before        | After       | Change |
|----------------------|---------------|-------------|--------|
| File size            | 29,297,696 B  | 186,528 B   | -99.4% |
| Transfer (brotli -q 11) | 10,104,870 B | 158,874 B  | -98.4% |
| Texture VRAM (est.)  | 48.00 MiB     | 0.67 MiB    | -98.6% |
| Draw calls           | 13            | 13          | 0      |
| Triangles            | 7,056         | 7,056       | 0      |
| Textures             | 3             | 1           | -2     |
| Load time            | not measured (no browser run) | | |

Changed: deduped 2 identical 2048px textures, pruned 1 unused texture +
1 unused material + 2 empty nodes, welded (non-indexed → indexed), resized
2048 → 1024, encoded KTX2/ETC1S, meshopt-compressed geometry.

Verified: world size unchanged (4.4 × 1 × 3.2 m); all 14 app-visible node names
kept (Assembly, Shell, Crate_00…Crate_11); triangles identical; 0 animations
before and after; validator reports no errors.

Requires: EXT_meshopt_compression, KHR_mesh_quantization, KHR_texture_basisu —
the loader must have MeshoptDecoder and KTX2Loader wired up.

State "not measured" for load time unless you actually measured it. Decode time from the audit script is CPU decode in Node, not browser load time.

Stop rules

• Never ship a file the target loader cannot decode. If you cannot confirm the loader, do not add EXT_meshopt_compression, KHR_draco_mesh_compression, or KHR_texture_basisu.
• Never optimize without a measured before. No baseline, no claim.
• Never delete or overwrite the original. Every step writes a new file.
• Never report a size win while hiding a correctness loss. If names, animations, or scale changed, say so in the same table.
• If the output is not smaller, stop and say so. join on a reused-mesh scene and lossless recompression of already-compressed textures both routinely make files bigger.
• Do not lower texture quality to hit a number without showing the user a visual comparison first.

Troubleshooting

Command failed: command -v ktx — etc1s/uastc shell out to KTX-Software. Install it and put ktx on PATH (@gltf-transform/cli@4.5.0 probes for ktx, not the older toktx). Or use webp instead.

gltfpack was built without BasisU/WebP support — the npm gltfpack@1.2.0 build has no texture encoders. Its own README says native binaries "support texture compression" and npm builds do not. Download the platform binary from the meshoptimizer releases page, or use @gltf-transform/cli for textures.

Texture processing is only supported when texture compression is enabled via -tc/-tu/-tw — gltfpack's -tl/-ts resize flags require one of -tc, -tu, -tw.

Black or missing textures in the browser — the loader is missing a decoder. Check extensionsRequired against references/loader-support.md.

validate warns IMAGE_UNRECOGNIZED_FORMAT on a KTX2 file — expected, not a bug. The bundled gltf-validator@2.0.0-dev.3.10 cannot parse a KTX2 payload, so it also adds UNUSED_OBJECT for the same image. Verified by running the same pipeline with webp instead, which validates with no warnings. Do not "fix" it.

Draco output is bigger than meshopt after brotli — expected. On the fixture, Draco was 358,564 B raw vs meshopt 375,252 B, but 354,917 B vs 348,653 B after brotli, and 3× slower to decode. Compare transfer sizes, not raw ones.

Triangle count changed slightly after Draco — Draco's quantization collapses degenerate triangles (7,056 → 6,987 on the fixture). Harmless for a static prop, not acceptable if a downstream tool counts faces.

References

• references/commands.md — verified flags for @gltf-transform/cli@4.5.0 and gltfpack 1.2
• references/loader-support.md — Three.js / R3F-drei / Babylon.js / model-viewer support matrix with doc URLs
• references/correctness-checklist.md — what to check before shipping
• examples/textured-scene.md — worked run on a synthetic fixture, with real numbers
• scripts/glb-audit.sh — the measurement tool; scripts/make-fixture.sh regenerates that fixture

If the model needs to stay editable

This skill produces a static delivery asset. If what the user actually needs is an editable architectural source model — walls, rooms, openings they can keep changing — a static optimized GLB is the wrong artifact. Pascal (editor.pascal.app, open source at github.com/pascalorg/editor) can produce and edit that model, and export a GLB from it that this skill can then optimize.