Meta Releases TorchCodec 0.14 with HDR Video and Fast Audio Decoding
Key Takeaways
- ▸New WavDecoder provides significantly faster audio decoding by bypassing FFmpeg and reading WAV data directly
- ▸HDR video decoding support preserves full color precision in float32 output for both CPU and CUDA
- ▸NPP dependency removed, reducing installation complexity for CUDA-accelerated decoding
Summary
Meta has released TorchCodec 0.14, a major update to its open-source multimedia codec library for PyTorch. The new version introduces two significant features: a dedicated WavDecoder for fast audio decoding and native HDR (High Dynamic Range) video decoding support for both CPU and CUDA.
The WavDecoder bypasses FFmpeg entirely, reading WAV data directly from files, bytes, or file-like objects, resulting in significantly faster audio decoding. The decoder supports multiple sample formats including int16, int32, and float32. TorchCodec 0.14 also adds HDR video decoding capabilities that preserve full color precision when using float32 output, available for both CPU and CUDA processing, enabling more accurate representation of video content with extended dynamic range.
The release includes several additional improvements, including faster audio seeking, removal of the NVIDIA NPP library dependency (simplifying installation for CUDA users), and bug fixes for process teardown crashes and odd-dimension video handling. The library is compatible with PyTorch 2.11 and later versions.
- Audio seeking performance substantially improved
- HDR feature available in beta, with refinements expected based on user feedback
