Credits and acknowledgements¶

VMAFx stands on other people's work. This page names every third-party item the project ships, vendors, adapts, learns from or uses, says how it reaches the project, and states the licence the upstream gives it. If you want to know who to thank, what a vendored file's terms are, or where a model's weights came from, start here.

The licence texts of the code that ships are in LICENSES/; the terms per file are in REUSE.toml and explained in Licensing. VMAFx itself is a fork of Netflix VMAF; the history is in the references.

At a glance¶

226 entries. By kind: upstream 2, code 22, library 67, tool 45, action 33, image 7, font 2, model 7, dataset 11, paper 15, standard 7, text 8. By relation: shipped 44, vendored 20, adapted 22, inspired 13, used-by-CI 72, integrated 55.

30 entries have a licence this page could not verify upstream: Bampis, Gupta, Soundararajan and Bovik, SpEED-QA (IEEE SPL 2017), Contributor Covenant 2.1, Debian base image, Developer Certificate of Origin 1.1, Documenting Architecture Decisions (Michael Nygard), Doxygen, Go base image, Hadolint, Intel oneAPI Base Toolkit, Intel oneAPI Base Toolkit image, ITU-R BT.2022, general viewing conditions for subjective assessment, ITU-R BT.2100, image parameter values for HDR television, ITU-R BT.2124-0 (2019), potential visibility of colour differences in television, Li, Zhang, Ma and Ngan, Image quality assessment by separately evaluating detail losses and additive impairments (IEEE TMM 2011), Mantiuk and Azimi, PU21 (QoMEX and PCS 2021), MATLAB reference implementations (iCID, SpEED, STMAD, matlabPyrTools, colorspace), matplotlib (Python), Mittal, Moorthy and Bovik, No-reference image quality assessment in the spatial domain (IEEE TIP 2012), Mittal, Soundararajan and Bovik, Making a completely blind image quality analyzer (IEEE SPL 2013), NASA/JPL Power of 10 and the JPL C coding standard, Python base image, QEMU, ROCm development image, SEI CERT C Coding Standard, Sheikh and Bovik, Image information and visual quality (IEEE TIP 2006), SMPTE ST 2084:2014, Perceptual Quantizer EOTF, Swatinem/rust-cache, Toward a better quality metric for the video community (Netflix, 2020), Toward a practical perceptual video quality metric (Netflix, 2016), Ubuntu base image.

How to read the tables¶

Each row is one item. The relation says how much of it reaches the project:

The licence is an SPDX expression exactly as the upstream states it, or one of three words: proprietary for a closed product, a service or custom terms used as their provider states them (the note says which), none when the upstream states no licence, and unknown when it could not be verified. An unknown is a gap to close, not a clearance: do not rely on it. GitHub-hosted licences were read from each repository on 2026-10-08.

Upstream projects and vendored or adapted code¶

Code that came from somewhere else. Files that carry a third-party copyright header keep the upstream terms (ADR-1250); the paths column lists them.

Libraries¶

Dependencies that are compiled into the binaries (Go modules) or imported by the Python tools. Only direct dependencies are listed; the release images carry the full notices (Licensing).

Models¶

Datasets¶

Training and evaluation data. Several datasets forbid redistribution or commercial use; VMAFx ships no dataset file. The terms are quoted from the dataset pages in Training data.

Papers and standards¶

Methods that VMAFx implements from the published description. Where an author's code exists, it is listed above under its own licence. The note of each row names the repository page that cites it.

Adapted texts and inspirations¶

Build and CI tools, actions and images¶

Fonts¶

Add an entry¶

You added, copied or learned from something that is not VMAFx's own work. To credit it:

1. Open docs/credits.yaml and add one entry. The fields are id (lower-case kebab), name, url (the upstream, https), kind (upstream, code, library, tool, action, image, font, model, dataset, paper, standard or text), relation (the table above), license, and optionally license_note, paths (repository paths or globs that use it), evidence (repository files that state the facts you wrote, required in spirit for models, datasets and papers) and note.
2. Copy the licence from the upstream. If it names none, write none; if you cannot tell, write unknown. Never guess.
3. Run make docs-fragments-write to render the tables, then make docs-fragments-check.

make docs-fragments-check runs scripts/docs/generate-credits.py --check and scripts/docs/check-credits.py. The second fails when:

• the tables differ from the list (page drift);
• a vendored or inherited third-party path has no entry: any non-project licence in REUSE.toml, any third_party, 3rdparty or vendor directory, notice and licence files, font files, and source files whose copyright header names someone other than the project and its contributors;
• a LICENSES/*.txt text is used by no entry and no project code;
• a skill or agent file says it was derived from an upstream (derived_from, or "Adapted from name") and no entry has that URL, which is how files that praetorctl adopt installs are held to their credit;
• an entry names a path or an evidence file that is not in the checkout.

Each check has a planted-defect test in scripts/docs/tests/test_credits.py that the gate must fail on. The decision is ADR-2485.