Key decisions: `#![forbid(unsafe_code)]` workspace-wide, patent-free codecs only (AV1/VP9/Opus/FLAC -- no H.264/H.265/AAC ever), async on Tokio, zero C/Fortran deps in default features, native WASM target.
This is v0.1.0 -- APIs are stabilized but not yet battle-tested at scale. Performance benchmarks vs FFmpeg/rav1e/dav1d coming soon.
Feedback on API design welcome, especially the filter graph and transcoding pipeline.
The mass of it is incredible, but it's a bit hard to tell what is an AI hallucination.
Digging in at random, it appears real, but it doesn't smell right for code of this mass. Most of the code seems to exist as it describes - which is an insane quantity. Everything seems to be what you would get if you asked AI to write it - for example, I looked through tonemapping and it appears to implement a small set of textbook tonemapping algos. In code that is really used, I expect to see something more purpose driven, not sure how to better say it. But an AI when asked to "handle tonemapping" would just pick a few literature methods and implement.
One random pick was "oximedia-gaming" which it said was stable with nvenc support. I wanted to check this out because I've called nvenc from C++ before and it was hard. The nvenc support is a no-op. So that's not quite right.
If (and it's a huge if), this code really works as it says, and it's written mostly by LLMs (it appears to be), then this is a huge testament to the value highly structured environment like Rust provides to the LLM.
I shudder to imagine the token costs...
Also, since it mentions full WASM support, a web demo would be nice.
WTF?