Releases · Gjusev/vision-input-check

First release of vision-input-check.

Regression tests for images sent to vision LLM APIs:

pixel-identical variants (original, png-reencode, exif-strip) with real pixel verification
lossy stress variants (jpeg-75, jpeg-40, resize-50, grayscale, webp)
N-run scoring with flaky exclusion and published denominators
strict CI gates: identical_delta_min, lossy_accuracy_min, max_flaky
offline ReplayProvider; OpenAI-compatible LiveProvider (extra [live])
standalone HTML report, bundled JSON Schema (schema_version 1.0)
synthetic offline demo with seeded failures (exit 1 strict / 0 relaxed)

pip install vision-input-check
vision-check demo --max-identical-delta 0.17 --min-lossy-accuracy 0.8

Navigation Menu