TorchCodec v0.14: High-Performance HDR Video and Rapid Wav Decoding

Released on June 3rd by NicolasHug, the v0.14.0 update to torchcodec introduces significant enhancements to how audio and video are handled, specifically targeting speed and color precision.

NicolasHug

🚀 Core Feature Highlights

The latest release focuses on two primary additions: a streamlined audio decoder and expanded capabilities for high-dynamic-range video.

1. Accelerated Wav Decoding

The new WavDecoder is designed for maximum efficiency. Unlike previous methods, it ~~relies on FFmpeg~~ completely bypasses FFmpeg, reading WAV data directly from the source.

Versatility: It can process data from files, raw bytes, or file-like objects.
Format Support: Compatible with various sample formats, including int16, int32, and float32.

Implementation Example:

from torchcodec.decoders import WavDecoder

# Initialize the decoder with a wav file
decoder = WavDecoder("audio.wav")
# Extract all samples (returns AudioSamples object with data and sample_rate)
samples = decoder.get_all_samples()

2. HDR Video Support

The VideoDecoder has been upgraded to support High Dynamic Range (HDR) video. This ensures that the rich color detail of HDR content is not lost during the decoding process.

To maintain full precision, users must set the output_dtype to torch.float32. This results in RGB frames mapped to the mathematical range: $\text{Pixel Value} \in [0, 1]$

Implementation Example:

from torchcodec.decoders import VideoDecoder

# Load HDR video ensuring float32 precision
decoder = VideoDecoder("hdr_video.mp4", output_dtype=torch.float32)
# Access the first frame with full HDR precision
frame = decoder[0]

⚠️ Note: This HDR functionality is currently in the beta stage. Expect potential behavioral adjustments as the team incorporates user feedback.

🛠️ Technical Improvements & Bug Fixes

The update also streamlines the installation process and resolves critical stability issues.

System Changes

Category	Change	Impact
Audio	Optimized seeking in `AudioDecoder`	Significantly faster navigation (#1449)
Dependencies	~~NVIDIA NPP Library~~ $\rightarrow$ Removed	Simplified CUDA installation/usage

Resolved Issues

CUDA Stability: Fixed a rare crash that occurred during the process teardown phase of the CUDA decoder (#1441).
Dimension Handling: Resolved an issue where CUDA decoding failed for videos with odd-numbered dimensions (#1462).

📊 Summary Checklist & Workflow

Release Progress:

Implement WavDecoder (FFmpeg-free)
Enable HDR float32 decoding
Optimize AudioDecoder seeking
Remove NPP dependency
Patch CUDA odd-dimension bug

Decoding Logic Flow: