Lighter models. Same shape.
An in-browser glTF/GLB optimizer — DRACO geometry compression and WebP textures, processed entirely on your machine.

---

Drop a .glb (or fully-embedded .gltf) into the page and Slipstream analyzes it, re-encodes every texture to WebP, compresses the geometry with DRACO, and hands you back a smaller GLB — all inside a Web Worker in your browser tab. There is no server, no upload, no account.

Privacy is architectural

Your model never leaves your machine: file → ArrayBuffer → Web Worker → download blob. This codebase has no upload path, and the end-to-end suite asserts zero non-local network requests during a full optimize run on every build. All wasm (DRACO encoder/decoder, image codecs) is served from the app's own origin.

What it does

• Analysis report — vertex/primitive/texture counts, per-texture roles derived from how materials actually bind them (baseColor, normal, metallic-roughness, occlusion, emissive), and feature detection: animation, skinning, morph targets.
• Textures → WebP — real libwebp (via jSquash), with per-role fidelity defaults: color maps (sRGB) go lossy at high quality; data maps (normals, metallic-roughness, occlusion — linear) stay lossless unless you explicitly opt into the aggressive preset, because lossy WebP visibly corrupts normal data. Downscaling is Lanczos3 and never upscales.
• Geometry → DRACO — KHR_draco_mesh_compression with tuned quantization. Automatically skipped when the model has skinning or morph targets: DRACO doesn't cover morph deltas and its re-indexing risks joint corruption, so those models keep their geometry and still get the texture savings.
• Presets + per-texture overrides — maximum / balanced / aggressive globally; exclude, quality, and max-resolution per texture. The plan shown in the texture table is the exact plan the pipeline executes.
• Before/after compare — a slider-wipe stage over two synchronized <model-viewer> cameras, rendering the actual optimized bytes.
• Honest numbers — the results breakdown shows growth as well as savings; a category that got bigger says so in red, it is never hidden.

Try it with the bundled demo model: NASA/JPL-Caltech's Perseverance rover (CC0, 213k vertices, 24 textures — see public/demo/ATTRIBUTION.md).

What it deliberately doesn't do

• No mesh decimation — vertex positions survive (within DRACO quantization); triangle counts are preserved. "Same shape" is the contract, and a read-back test suite holds the optimizer to it.
• KTX2/Basis textures are kept as-is, not re-encoded.
• Files referencing external .bin/image resources are rejected with a clear message — a client-side tool can't fetch them, and guessing would be worse.

Stack

Astro shell with a single Preact island; the optimizer core is framework-agnostic TypeScript on glTF Transform, running in a Web Worker via Comlink; draco3d and jSquash wasm, self-hosted; <model-viewer> for the compare stage.

Development

npm install          # postinstall self-hosts the DRACO wasm into public/
npm run dev          # dev server on :4321
npm run typecheck && npm run lint && npm test    # strict TS · ESLint · vitest
npx playwright test  # e2e: full optimize runs in a real browser
npm run build

License

Code is MIT. The demo model is CC0 (NASA/JPL-Caltech). Fonts: Azeret Mono (OFL) is self-hosted; Clash Display (ITF FFL) and Mona Sans load from their own CDNs and are not redistributed in this repo.