Catetus

Production infrastructure for Gaussian Splats — compress, validate, and ship splat assets with standards-aligned output and reproducible quality gates.

Website	catetus.com — live demos, leaderboard, docs
Try it	Drop a .ply in the browser
Benchmark	SplatBench leaderboard — 28 scenes (14 real + 14 synthetic), open submission
Standards	KHR conformance report — 30 clauses, 13 fixtures

---

What is Catetus?

Gaussian Splatting is moving from research demos into production web, mobile, and spatial apps. Capture tools export huge .ply files; runtimes expect compact, standards-aligned assets; teams need proof that optimization did not destroy visual quality.

Catetus is the delivery layer in the middle — think FFmpeg + Lighthouse for Gaussian Splats:

• Ingest .ply, .spz, and glTF Gaussian Splat assets
• Optimize for real device byte and fidelity budgets (CLI presets + hosted API)
• Export glTF KHR_gaussian_splatting and SPZ — no proprietary container format
• Prove quality with deterministic visual diff, SplatBench, and conformance suites

The open-source core (this repo) ships the CLI, viewer SDK, benchmark corpus, and Khronos/OpenUSD conformance tooling. Hosted optimize, premium passes, and enterprise deployment live on catetus.com.

---

Why teams choose Catetus

Capability	What it means
Standards-first output	glTF KHR Gaussian Splatting + SPZ; assets work in mainstream viewers and DCC pipelines
Reproducible pipeline	Same input + preset → byte-identical output and stable BLAKE3 scene hashes
Public benchmark	SplatBench — 28 scenes (14 real + 14 synthetic stress tests), open encoder comparison
Quality gates	Per-scene fidelity (ΔE94, SSIM), visual diff harness, KHR conformance crate in CI
Ship anywhere	Rust CLI, @catetus/viewer (WebGPU + WebGL2), GitHub Action, REST API

Canonical-11 leaderboard (Mip-NeRF 360 + Tanks-and-Temples + Deep Blending, wmv-vq45-no-prune-tight preset):

Metric	Value
Scenes measured	11 / 11
Mean compression vs input PLY	19.8× (range 16.6× – 21.9×)
Mean PSNR (gsplat, 512², orbit-8, SH=3)	47.45 dB (median 47.78, min 43.46)
Mean SSIM	0.9991 (min 0.9973)

Per-scene table: canonical-11.md · machine-readable: canonical-11.json

Broader SplatBench v0 corpus (28 scenes including synthetic stress probes, web-mobile preset):

Metric	Value
Median compression (28 corpus scenes)	22.7×
Real outdoor (bicycle, 3.6M splats)	25.5× (855 MB → 34 MB)
Real indoor (bonsai, 1.2M splats)	22.8× (274 MB → 12 MB) 1
Fidelity gates passing	16 / 16 fidelity-gated scenes (web-mobile + size-min both pass)

Coverage scope: 16 of the 28 corpus scenes have fidelity oracles registered (bonsai + bicycle real photogrammetry + 14 synthetic stress probes). The remaining 12 are size-benchmarked only: 6 await scaffold-GS render oracles, 5 are LOD-ladder synthetic without ground-truth captures (cluster_fly_*), 1 is an under-trained iter7k variant. Coverage expansion tracked on the SplatBench leaderboard page.

Full tables and per-scene breakdown: leaderboard · report · interactive HTML

---

Quick start

Prerequisites: Rust stable (≥ 1.74), Node.js 20+, pnpm 9+. See INSTALL.md for platform notes.

git clone https://github.com/Catetus/catetus.git
cd Catetus
./scripts/install-githooks.sh   # one-time: enables the partnership-docs pre-push guard (see docs/CONTRIBUTING.md)
./setup.sh

# Inspect a splat
./target/release/catetus analyze fixtures/tiny/basic_binary.ply --pretty

# Optimize for web delivery (glTF + chunked buffers)
./target/release/catetus optimize fixtures/tiny/basic_binary.ply \
  --preset web-mobile --out /tmp/scene.gltf

# Preview in the browser (WebGPU viewer)
./target/release/catetus preview /tmp/scene.gltf

Run tests: make test · Run SplatBench locally: make bench-splatbench

Step-by-step guide: docs/getting-started.md

---

CLI commands

Command	Description
analyze	Deterministic JSON report (size, bounds, attribute stats)
inspect	Validate an asset; print a short summary
convert	Convert between PLY, SPZ, glTF, GLB
optimize	Run a preset (web-mobile, size-min, geospatial, …)
preview	Local WebGPU viewer
diff	Before/after visual diff report (Playwright-backed)
benchmark	Device-profile timings
corpus run	Run a named SplatBench suite
submit	Submit a job to the hosted API
spec-check	Validate against KHR / extension rules

catetus optimize scene.ply --preset web-mobile --out out/
catetus diff scene.ply out/scene.gltf --threshold 0.03 --out reports/diff/

---

Integrations

Surface	Link
GitHub Action	apps/optimize-action — PR gate on compression + fidelity badge
Hosted API	https://api.catetus.com — job create, upload, status, fidelity scoring (server source is private; see OpenAPI)
Viewer SDK	@catetus/viewer — WebGPU compute decode + streaming LOD
Blender add-on	integrations/blender

---

Repository layout

Catetus/
  crates/
    catetus-core/          # SplatIR — canonical internal representation
    catetus-ply/           # PLY ingest + write
    catetus-spz/           # SPZ I/O
    catetus-gltf/          # glTF KHR Gaussian Splatting + GLB
    catetus-optimize/      # Optimization pass framework
    catetus-khr-conformance/  # KHR extension validator (30 clauses)
    catetus-usd/           # OpenUSD writer (draft)
    catetus-cli/           # `catetus` binary
  packages/
    viewer/                   # @catetus/viewer
    report-ui/                # Diff + parity HTML reports
  specs/                      # Feature specs (SPEC-0001 … SPEC-0013)
  benches/                    # SplatBench corpus + reports
  apps/
    web/                      # Marketing site + leaderboard (Astro)
    api/                      # Hosted optimize API (Rust / Axum)
    optimize-action/          # GitHub Action
  docs/                       # User + architecture docs
  tests/                      # Integration + visual regression

Architecture overview: docs/architecture.md

---

Standards & conformance

Catetus targets glTF 2.0 + KHR_gaussian_splatting as the primary interchange format, with SPZ as a first-class compressed wire format. Advanced streaming uses a vendor extension (CT_spatial_streaming_index) that degrades gracefully in viewers that ignore it.

• KHR suite: cargo test -p catetus-khr-conformance — conformance report · live matrix
• KHR SPZ compression extension (draft): docs/standards/KHR_gaussian_splatting_compression_spz.md
• OpenUSD: writer + conformance work in progress (SPEC-0011, SPEC-0012)

---

Documentation

Doc	Audience
INSTALL.md	First-time setup (macOS, Linux, Windows)
docs/getting-started.md	End-to-end CLI walkthrough
docs/architecture.md	System design
specs/	Feature specifications
CHANGELOG.md	Release history
CONTRIBUTING.md	How to contribute

---

Contributing

Issues and PRs welcome. We are spec-driven and determinism is non-negotiable — see CONTRIBUTING.md for the workflow, DCO, and review expectations.

• Report a bug
• Request a feature
• Submit a scene to SplatBench

---

License

Apache-2.0. See LICENSE.

Footnotes

1. The two bonsai rows above measure two different files at two different presets — they are not directly comparable. SplatBench v0 uses a 274 MB iter7k bonsai derivative at the bytes-first web-mobile preset; the canonical-11 row uses the official Inria iter30k bonsai (308.7 MB, md5 ad5377eb…) at the fidelity-first wmv-vq45-no-prune-tight preset. Identical scene names, different inputs and budgets. ↩