Rust implementation

[Filyus]

Hello everyone! It took me about a month of vibe coding in my spare time to create the Draco library entirely in Rust. It wasn't easy and required a lot of patience, but I'm finally happy to share the code with everyone:
https://github.com/Filyus/draco

First, I split the C++ code into two projects and moved all the .cc files to root folders:

• https://github.com/Filyus/draco/tree/main/draco_core
• https://github.com/Filyus/draco/tree/main/draco_io

Then I created the code in Rust:
https://github.com/Filyus/draco/tree/main/crates
I also split it into draco-core and draco-io, which supports converting geometry (and, if possible, normals) to and from GLTF/GLB, FBX, OBJ, and PLY.

I also wrote WASM modules and a sample web application for conversion:
https://github.com/Filyus/draco/tree/main/web

As a result, gltf_reader.wasm is 203 KB (~245 KB now). For context, Google's draco_decoder.wasm is ~286 KB, and the glTF-oriented draco_decoder_gltf.wasm is ~193 KB. Unlike those standalone Draco decoders, this module also parses GLTF/GLB files.

The code now passes the Rust workspace tests and web workspace tests, and the branch has been rebased onto the current google/main. I'm continuing with final cleanup and hardening before preparing it as a PR.

[Filyus]

I’ve implemented all prediction schemes and full compression level support (with some assistance from an LLM). I’m planning to release it today or tomorrow.

If my measurements are correct, the results look very promising: the Rust implementation is ~10× faster for encoding and ~5× faster for decoding, while producing byte-for-byte identical .drc output.

Stay tuned. Thanks for keeping this issue open.

[jfuruness]

any updates on this out of curiosity?

[Filyus]

Sorry for the delay — ended up being a month rather than a day!

The implementation is now feature-complete: all prediction schemes (including UV TexCoords), Edgebreaker with valence-based traversal, all attribute types, zero unsafe code, and byte-for-byte identical output to the C++ reference encoder across all speeds 0–10.

Performance: Rust is ~2× faster on encode and ~20–30× faster on decode for speeds 0–9 (Edgebreaker). Speed 10 (sequential / no Edgebreaker) is essentially at parity now.

The changes are now pushed. Thanks for your patience!

There may be some inaccuracies in the tests or bugs in the code, but I'm still improving it.

[jfuruness]

Very cool! Hopefully it ends up getting picked up by google. I didn't end up using it for the final product (I didn't realize texture and uv compression wasn't lossless, and I'm not all too familiar with this field) but this is a really cool project!

[Filyus]

Quick update: the current Rust implementation is here:
https://github.com/Filyus/draco

Since the original post, the Rust implementation has become much more complete and better tested:

• Full Rust workspaces for draco-core, draco-io, draco-cpp-test-bridge, and the WASM/web modules.
• Mesh and point-cloud encode/decode paths, including sequential and EdgeBreaker coverage.
• C++ interop tests in both directions: C++ encode -> Rust decode and Rust encode -> C++ decode.
• Legacy .drc fixture coverage, malformed input hardening, and real-file decode comparisons.
• Current test pass is green for both workspaces:
  • cargo test --manifest-path crates/Cargo.toml --all
  • cargo test --manifest-path web/Cargo.toml --all

Generated mesh encode/decode benchmark

Mesh	Speed	Faces	C++ encode	Rust encode	Enc result	C++ decode	Rust decode	Dec result
sphere 24x48	0	2208	4003 us	708.4 us	Rust 5.65x faster	9351 us	186.0 us	Rust 50.27x faster
sphere 24x48	1	2208	3849 us	636.4 us	Rust 6.05x faster	9230 us	159.8 us	Rust 57.75x faster
sphere 24x48	2	2208	3379 us	283.8 us	Rust 11.91x faster	9261 us	123.9 us	Rust 74.72x faster
sphere 24x48	3	2208	3424 us	301.1 us	Rust 11.37x faster	9113 us	101.2 us	Rust 90.05x faster
sphere 24x48	4	2208	3388 us	252.1 us	Rust 13.44x faster	9165 us	107.6 us	Rust 85.17x faster
sphere 24x48	5	2208	3468 us	225.1 us	Rust 15.41x faster	9188 us	89.6 us	Rust 102.57x faster
sphere 24x48	6	2208	3441 us	226.7 us	Rust 15.18x faster	9160 us	76.3 us	Rust 120.11x faster
sphere 24x48	7	2208	3511 us	225.3 us	Rust 15.58x faster	9085 us	78.8 us	Rust 115.27x faster
sphere 24x48	8	2208	3252 us	219.3 us	Rust 14.83x faster	9161 us	72.6 us	Rust 126.26x faster
sphere 24x48	9	2208	3294 us	213.1 us	Rust 15.46x faster	9171 us	72.4 us	Rust 126.63x faster
sphere 24x48	10	2208	92 us	54.2 us	Rust 1.70x faster	33 us	29.5 us	Rust 1.12x faster
cube subdiv20	0	4800	9281 us	1439.8 us	Rust 6.45x faster	20896 us	317.1 us	Rust 65.89x faster
cube subdiv20	1	4800	9277 us	1432.6 us	Rust 6.48x faster	20806 us	309.9 us	Rust 67.13x faster
cube subdiv20	2	4800	8111 us	587.2 us	Rust 13.81x faster	20629 us	212.1 us	Rust 97.25x faster
cube subdiv20	3	4800	8045 us	603.2 us	Rust 13.34x faster	20689 us	208.5 us	Rust 99.21x faster
cube subdiv20	4	4800	8149 us	602.1 us	Rust 13.53x faster	20710 us	202.6 us	Rust 102.21x faster
cube subdiv20	5	4800	7957 us	574.0 us	Rust 13.86x faster	20362 us	179.5 us	Rust 113.45x faster
cube subdiv20	6	4800	8056 us	529.3 us	Rust 15.22x faster	20573 us	181.0 us	Rust 113.65x faster
cube subdiv20	7	4800	8264 us	526.1 us	Rust 15.71x faster	20561 us	184.2 us	Rust 111.63x faster
cube subdiv20	8	4800	7950 us	499.0 us	Rust 15.93x faster	20464 us	167.6 us	Rust 122.08x faster
cube subdiv20	9	4800	7871 us	502.7 us	Rust 15.66x faster	20568 us	162.4 us	Rust 126.64x faster
cube subdiv20	10	4800	188 us	106.3 us	Rust 1.77x faster	72 us	61.8 us	Rust 1.17x faster