Visual Verifier

Anonymization QA for images and videos

Your anonymizer says the video passed. Frame 4 still exposes the licence plate.

Visual Verifier tests blur, redaction, masking, and other visual processing frame by frame, and returns a deterministic PASS/FAIL, temporal evidence, and CI-ready reports.

Visual Verifier brings software-testing discipline to processed image and video output. Instead of trusting that an anonymization or visual-processing job completed, it independently checks what happened frame by frame, verifies required regions when targets are supplied, and produces reproducible PASS/FAIL evidence for CI and review.

pip install visual-verifier
visual-verifier demo

FAIL
Frames missed: 4, 8, 12

No clone, no fixtures to download, no media of your own required. The sample is generated on your machine.

Runs entirely on your machine

Your media is never uploaded. The package imports no networking library at all, and a test enforces that on every commit. This documentation site loads no third-party fonts or trackers either.

Anonymization is the named use case, not the limit

The engine is a general pixel-level change verifier. It measures what actually changed between a reference and a candidate, region by region and frame by frame, so anything that alters pixels can be verified the same way.

Use it for	What a FAIL means
Blur, pixelation, redaction, masking	A frame was left unprotected
Watermark and overlay application	The mark is missing or too faint
Encoding, transcoding, filter chains	A stage silently did nothing
Compositing, inpainting, style transfer	The edit did not land everywhere
Any processed-media regression test	Output drifted from the reference

Privacy work is where the failure is most expensive, so it is what the documentation leads with. The verification contract itself makes no assumption about why the pixels changed.

The missing test after visual processing

Image and video pipelines are usually tested as software: the function ran, the model returned detections, the encoder produced a file, and the job exited successfully.

None of that proves the output media is correct.

An anonymizer can complete successfully while missing one frame. A masking pipeline can modify the frame while missing the required face or licence plate. A watermark job can finish while the mark disappears during part of a video. In every case the exit code is zero and the media is wrong.

Visual Verifier tests the produced media itself. It is not an anonymizer; it is the independent test that runs after one:

any anonymizer, redactor, or image/video processor
                      ↓
               processed output
                      ↓
              VISUAL VERIFIER
                      ↓
        Did the transformation happen?
        Did it happen in every frame?
        Did it cover the required target?
        Where exactly did it fail?
                      ↓
           PASS / FAIL + evidence
                      ↓
             fix → rerun → verify

Instead of asking only whether the software executed, it asks whether the expected visual transformation actually appeared, frame by frame, and — when reviewed targets are supplied — in the required region.

Think of it as regression testing for processed visual media: deterministic PASS/FAIL, the exact failing frames and regions, reviewable evidence, and CI-ready exit codes.

Your code has tests. Why shouldn't your processed media?

That also means tools like face blurrers, plate redactors, and OCR scrubbers are not competitors. They are upstream systems this verifies.

Start here

Document	Read it when you want to
Getting started	Install it and see a real failure in 60 seconds
Anonymization QA	Understand exactly what a PASS does and does not prove
Use it in CI	Fail a build on an anonymization gap
Limitations	Know the failure modes before trusting a PASS

Reference

Document	Contents
Command line	Every command, option, and exit code
Python API	Functions, result models, and error codes
Output schema	Field-by-field meaning of every report
Temporal tracking	Association, lifecycle, lineage, and metrics
Verification contract	The formal inputs and outputs
Target annotation	Reviewed target schema and coverage behaviour
Benchmark	Measured comparison against global-metric baselines
GitHub Action	Every action input, output, and permission

Scope and roadmap

Document	Contents
Problem statement	Why this exists and what it excludes by design
Use cases	Supported uses and unsupportable claims
Policy system	The planned V5.4 selectable policy protocol

Working on the code

Document	Contents
Architecture	Module boundaries and dependency direction
Code style	Naming, typing, docstring, and length rules
Release checklist	Everything a release must satisfy
Development history	How the prototype became this package

Project files live in the repository: AGENTS.md, CONTRIBUTING.md, CHANGELOG.md, ROADMAP.md, SECURITY.md, and CODE_OF_CONDUCT.md.