For three.js / WebGL work, glTF size and texture VRAM often bite before the first frame feels right. I used to guess how hard to simplify, re-export, and repeat. I wanted the data first.
So I made Glint: https://glintlab.cn
(Backup if the main site is down: https://glint-bg3.pages.dev/)
Open it in a desktop browser — nothing to install. It accepts .glb / .gltf only, hard cap 500MB per file (you get a warning past ~100MB).
Drop a file in, orbit it. You get a breakdown: triangles, verts, meshes/materials/textures, a VRAM estimate, risk flags — parsed in the browser with gltf-transform. That breakdown can feed an AI analysis (only the report text is sent, not the raw model; you need your own API key — without a key, AI analysis doesn’t run). The breakdown and local optimize don’t need a key.
After the breakdown (or the AI analysis report), optimize with presets or pick steps yourself — dedupe, weld, simplify, texture resize, Draco, Meshopt, KTX2 via toKtx (JPEG/PNG → KTX2/BasisU; viewer uses KTX2Loader). Some steps are lossy; the UI shows that. When you’re done, compare size / tris / VRAM estimate and download the new .glb.
It’s gltf-transform plus a readable report: understand the data, then optimize. KTX2 is supported. Not a magic one-click compressor, and I didn’t invent a new codec.
If you load models with three.js (Draco / Meshopt / KTX2 paths included), try a real file and tell me what breaks — browser, which step, what kind of model.
