ANSNR (removed)¶
This metric was removed
The float_ansnr feature extractor was removed from every backend in PR #38 (see ADR-0709 and the follow-up ADR-0749). ANSNR is a pre-VMAF metric (2001) that Netflix never adopted in production; it was carried by the upstream codebase but no shipped VMAF model references it. This page is retained for traceability.
Historical description¶
ANSNR (Average Noise-to-Signal Ratio) measured the ratio of distortion energy to signal energy averaged over the frame. It was registered as the float_ansnr feature extractor and emitted float_ansnr (luma) plus float_ansnr_cb / float_ansnr_cr when enable_chroma=true.
Migration¶
| Old surface | Status after PR #38 |
|---|---|
--feature float_ansnr on the CLI | Unknown feature; the CLI errors out. |
_METRIC_TO_EXTRACTOR["float_ansnr"] (Python helper in ai/data/feature_extractor.py) | Mapping entry removed; passing features=["float_ansnr"] raises. |
float_ansnr_hip / float_ansnr_cuda / float_ansnr_sycl (GPU twins) | Sources removed together with the CPU registration. |
VmafLegacyQualityRunner (Python runner that fed ANSNR into a libsvm model) | Sunset in ADR-0749. Use VmafQualityRunner against a modern .json model instead. |
Callers that need a per-frame distortion-energy signal should use PSNR (psnr_y / psnr_cb / psnr_cr) or PSNR-HVS (psnr_hvs); both remain first-class extractors with active CPU + GPU paths.