v0.7.0

0.7.0 (2026-09-29)

Features

• choose how frames are resized with interpolation (#15) (6317322)

v0.6.0

0.6.0 (2026-09-20)

Features

• read approximate frames (#11) (2158899)

v0.5.0

0.5.0 (2026-09-20)

⚠ BREAKING CHANGES

• read_all is removed; use list(iterframes.read(...)).

Features

• build wheels for Windows (#4) (39702a5)
• read frames by number, in the style of decord (#10) (6d0627e)
• read frames in batches with read_batches (#6) (136ae14)

v0.4.0

0.4.0 (2026-09-19)

The rewrite: static FFmpeg wheels, frames without copies, and hardware
decoding.

Breaking

• Python 3.11 or later is required, for the buffer protocol in the
  stable ABI
• read_batch is replaced by read_all, which returns NumPy arrays in
  the order of the video. read_batch returned Frame objects in reverse
  order
• Frames are C-contiguous arrays instead of strided views on a buffer
  padded at the end of each row
• FrameReader yields Frame objects, which support the buffer
  protocol, instead of (bytearray, height, width) tuples

Added

• The wheels bundle a static FFmpeg 9.0, so FFmpeg no longer needs to be
  installed. One abi3 wheel per platform covers every CPython from 3.11
  on: Linux x86_64 and aarch64 (manylinux_2_28), macOS arm64
• AV1 decoding, through dav1d
• Errors are raised instead of ending the iteration silently:
  FileNotFoundError and other OSErrors when the file cannot be opened,
  ValueError when it is not a video, RuntimeError when decoding fails
• read accepts os.PathLike paths
• FFMPEG_VERSION, the version of the linked FFmpeg
• Hardware decoding with device, named as in PyTorch: "mps" for
  VideoToolbox on macOS, "cuda" for NVDEC on Linux, which also resizes
  on the GPU, or "auto". DEVICES lists the devices of the build
• on_device=True, with device="cuda", leaves the frames on the GPU as
  CudaFrame objects in NV12, whose planes support DLPack
• Documentation site on Read the Docs

Changed

• FFmpeg decodes on several threads, and waiting for a frame releases the
  GIL, so other Python threads keep running
• The conversion to RGB, and the resizing, run on several threads too
• Frames reach NumPy without a copy: the arrays share memory with the
  FFmpeg frames
• The decoder thread stops as soon as the iterator is dropped
• A change of resolution or pixel format within a stream no longer stops
  the decoding
• Development uses uv; requirements-dev.txt,
  the Dockerfile, and the build scripts are replaced by uv.lock and
  scripts/build-ffmpeg.sh
• Tests compare the frames with PyAV instead of decord
• CI lints Rust and Python, tests every wheel, and publishes to PyPI
  through trusted publishing
• Updated to PyO3 0.29 and Rust edition 2024
• ffmpeg-next and ffmpeg-sys-next are replaced by build.rs, which
  builds the static FFmpeg, links it, and generates its bindings, so that
  maturin build needs no FFmpeg installed and no environment variables.
  The system FFmpeg is no longer used, and the static feature is gone