avtensor is a Rust-based library with Python bindings that decodes video, audio, and images directly into PyTorch tensors via FFmpeg, optimized for ML data-loading pipelines.

Stars

61

7-day growth

No data

Forks

1

Open issues

0

License

Apache-2.0

Last updated

2026-07-27

AI repository intelligence
FR-AI / ANALYSIS

Why it is worth attention

It offers native support for cloud storage inputs (gs://, s3://, HTTP(S)), combined video+audio decoding in one pass, automatic HDR-to-SDR tone mapping, and significantly faster decode-time resize compared to torchcodec.

Who it is for

  • PyTorch users building ML data pipelines
  • Data scientists working with video/audio datasets
  • ML engineers needing efficient media decoding with GPU acceleration
  • Researchers processing large-scale multimedia collections

Use cases

  • Decoding video frames and audio samples for training or inference
  • Preprocessing media with resize, frame-rate resampling, and loudness normalization at decode time
  • Reading media from cloud storage without intermediate file downloads
  • GPU-accelerated decoding with NVDEC for reduced CPU load and faster throughput

Strengths

  • Native cloud input support (gs://, s3://) without extra dependencies
  • Single decode pass for both video and audio streams
  • Automatic HDR to SDR tone mapping preserving color accuracy
  • Decode-time resize 1.7–6× faster than torchcodec in benchmarks

Considerations

  • Requires FFmpeg >=7.1 shared library and build-time dependencies (Rust, libtorch)
  • NVDEC hardware decode has limited concurrent engines and does not support H.264 4:4:4 chroma
  • GPU-resident output disables frame-rate resampling and HDR tone mapping

README quick start

Installation

Description

High-performance media decoding straight into PyTorch tensors

Related repositories

Similar projects matched by category, topics, and programming language.

l0ng-ai
Featured
l0ng-ai GitHub avatar

tty7

tty7 is a fast, GPU-accelerated terminal workbench with persistent sessions, built-in input tools, SSH support, and coding agent awareness.

Developer ToolsCLI & Terminal
359
steelbrain
Featured
steelbrain GitHub avatar

reims-vgpu

reims-vgpu is an experimental virtual GPU for macOS guests that uses QEMU to decode the guest's GPU command stream and execute it through Metal or Vulkan, leveraging the built-in AppleParavirtGPU driver without requiring custom kexts.

Desktop Apps
143
m-novotny
Featured
m-novotny GitHub avatar

memguard-rs

A Rust library that provides secure memory handling primitives including zeroization on drop, memory locking, constant-time comparison, and compile-time guarded regions, with zero dependencies and no_std support.

Embedded & IoTSecurity
131