Video Processing SciKit

Borrowing coding styles and conventions from scikit-image and scikit-learn, scikit-video is a Python module for video processing built on top of scipy, numpy, and FFmpeg.

This project is distributed under the 3-clause BSD.

Visit the documentation at http://www.scikit-video.org

Dependencies and Installation

Requirements:

• FFmpeg >= 2.8 on the system PATH (primary tested backend). The libav/avconv backend is deprecated and will be removed in a future release.
• Python >= 3.10
• numpy >= 1.22
• scipy >= 1.9
• Pillow >= 9.0
• mediainfo (optional; deprecated, will be removed in a future release)

Installation from PyPI:

pip install scikit-video

Installation from source (GitHub):

git clone https://github.com/scikit-video/scikit-video.git
cd scikit-video
pip install .

For development (editable install with test dependencies):

pip install -e ".[test]"

Known conflicts

If you installed scikit-video prior to version 1.1.10, you may have an import conflict with the older sk-video package. Run the following to fix it:

pip uninstall sk-video

Then check that skvideo resolves to the expected location:

import skvideo
print(skvideo.__file__)

What's new

scikit-video is actively maintained again.

• Deprecations and cleanup (1.3.0). The libav/avconv backend, mprobe/mediainfo, and the hardcoded container-extension allowlist are deprecated (DeprecationWarning; removal planned for 1.4.0) -- unknown extensions now warn and defer to ffmpeg instead of raising. Binaries are resolved only from absolute PATH entries (never the CWD). Roughly 1,000 lines of dead, vendored, and Python-2-era code removed, with a pyflakes lint gate in CI keeping it that way.
• Metric accuracy overhaul (1.2.0). The NIQE, BRISQUE, VIIDEO, and Video-BLIINDS implementations were validated against their reference MATLAB implementations and fixed where they diverged (reference NIQE model, antialiased half-scale resize, faithful VIIDEO port, motion tie-breaking). Breaking: these four metrics return different -- more accurate -- values than 1.1.x. Scores and features are not comparable across the 1.1.x/1.2.x boundary. Function signatures are unchanged. MSE/PSNR/SSIM/MS-SSIM and ST-RRED were validated as already correct and are unchanged.
• Maintenance (1.2.1). NumPy 2.5 compatibility (linalg.eig now always returns complex; symmetric eigendecompositions use eigh), the unused opencv-python-headless dependency removed, FFmpeg version/binary detection fixes, and clearer errors on bad backend= names and unsupported channel counts.
• The 1.1.12–1.1.15 line modernized the package for current Python (3.10+), NumPy 2.x, and SciPy with no breaking API changes: modern pyproject.toml packaging, non-local I/O (URLs and file-like objects), pathlib.Path support, audio passthrough (audiosrc=), a Python-3 correctness pass, and exception-based input validation.

See CHANGELOG.rst for the complete per-version history.

TODO/Roadmap

• Spatial-Temporal filtering helper functions
• Speedup routines (using cython and/or opencl)
• More ffmpeg interfacing
• Add additional algorithms and maintain more comprehensive benchmarks

For Contributors

See CONTRIBUTING.rst for setup instructions.

Testing

After installing with the [test] extra, run the suite with:

pytest -v skvideo/tests

Tests that require FFmpeg or libav are automatically skipped if the corresponding binary is not on the system PATH.

Copyright 2015-2026, scikit-video developers (BSD license).