Compress 3D Models

Duplicate meshes and unused data are removed, vertices are welded, attributes are quantized to integers and the buffers are packed with EXT_meshopt_compression. Optionally the triangle count is reduced and oversized textures are scaled down.

Cutting triangles works by collapsing edges, and that needs neighbouring triangles to share their vertices. STL and most CAD exports store a separate normal for every face, so no two triangles match up and there is nothing left to collapse. When that happens the original normals are dropped so the mesh can be welded, and flat ones are written again once the geometry is thinner. Models that already carry smooth shared normals keep them untouched.

Meshopt compressed GLB is read by three.js, Babylon.js, model-viewer, Godot and the Khronos glTF validator. If a tool refuses it, turn meshopt off and keep only quantization, which needs no extension support at all.

You can drop one in and it will be read into a GLB first, because the compression steps are glTF features. The download is not locked to GLB though: pick any format the converter writes and the optimized model is written out to it, keeping the reduced triangle count and the resized textures.

Because compressed data has to be described by the file format itself, and glTF is the only one here that standardises it. STL spends fifty plain bytes on every triangle with nowhere to record that they were packed, and OBJ is plain text. 3MF is the one other format that comes out smaller for free, since the file is a zipped package. Everywhere else the ways to save space are fewer triangles, binary rather than ASCII encoding, and fewer decimal places.

The Draco encoder is not available as a browser friendly WebAssembly module in this pipeline, and meshopt usually wins on decode speed anyway. Meshopt gives comparable size with far less CPU cost when the model loads. Reading is a different matter: Draco compressed files are unpacked without any extra steps, since the decoder ships with the site.