[00] · BENCHMARK

3DAssetBench. An honest baseline.

Real assets vendored from the Khronos glTF-Sample-Assets corpus, optimized through SpatialPack and a stock @gltf-transform/functions pipeline at matching versions. Δ is the SpatialPack output minus the gltf-transform output — negative means SpatialPack is smaller on that asset.

How this differs from /benchmark

Same corpus, different measurement — read them as two instruments, not two opinions.

This page — raw GLB bytes on disk, no gzip, against a gltf-transform baseline computed in-process from the same library version we link. It also carries the per-asset cohort tags and predicted frame cost that the analyzer's budget suggestions are derived from. It is the regression table.
/benchmark — gzip-9 bytes over the wire, against the real gltf-transform and gltfpack command-line binaries, with a rendered ΔE94 / SSIM quality gate and every loss published. It is the competitive leaderboard.

Compressing then gzipping is a smaller relative win than compressing alone, so the medians differ by design. Per asset the two agree: DamagedHelmet reads 89.6% on both.

Readiness — “will it open cleanly in AR” — is a third question on a third corpus, published at /ai3dbench.

Loading benchmark…