EncodingDB
GitHubRun a benchmark

PL Score v7

Methodology

PL Score v7 combines quality, bitrate efficiency, and encode speed into one fixed score for a single benchmark workload.

v7.0 · pilot

PL7=100×\mathrm{PL7} = 100 \times×\times×\times
Quality 50%Bitrate 30%Speed 20%

Quality

50%

Uses both average VMAF and the lower tail so short bad moments are not hidden by a strong mean.

0.85 Vmean+0.15 VP50.85\,V_{\mathrm{mean}} + 0.15\,V_{P5}

Calculation

Effective qualityVPL=0.85 Vmean+0.15 VP5V_{\mathrm{PL}} = 0.85\,V_{\mathrm{mean}} + 0.15\,V_{P5}
Quality termQ=[clamp⁡(VPL−2080,0,1)]2.4Q = \left[\operatorname{clamp}\left(\frac{V_{\mathrm{PL}} - 20}{80}, 0, 1\right)\right]^{2.4}
Bitrate termB=11+RencRrefB = \frac{1}{1 + \frac{R_{\mathrm{enc}}}{R_{\mathrm{ref}}}}
Speed termS=1−exp⁡ ⁣(−1.2 FPSencFPSsrc)1−exp⁡ ⁣(−4.8)S = \frac{1 - \exp\!\left(-1.2\,\frac{\mathrm{FPS}_{\mathrm{enc}}}{\mathrm{FPS}_{\mathrm{src}}}\right)}{1 - \exp\!\left(-4.8\right)}

General PL7

General PL7 requires complete, equal-class suite coverage.

100exp⁡ ⁣(1n∑i=1nln⁡ ⁣PL7i100)100\exp\!\left(\frac{1}{n}\sum_{i=1}^{n}\ln\!\frac{\mathrm{PL7}_i}{100}\right)

Rules

  • Missing required measurements produce no canonical PL7 score.
  • Bitrate is compared with a fixed reference for that workload, not the current leaderboard.
  • PL Score, PL Fit, and confidence are separate concepts.
  • Every score is tied to versioned benchmark and provenance data.
Technical notesprovenance, limitations, and score semantics

Required measurements: VMAF mean, VMAF P5, video bitrate, encode FPS, source FPS, and workload reference bitrate.

Provenance: formula version, benchmark protocol version, source suite version, workload ID, reference bitrate, and VMAF model ID.

PL7 describes a workload under a benchmark protocol. It is not a universal statement about an encoder.

PL Fit is the personalized ranking layer. Confidence describes evidence quality. Neither changes the public PL7 formula.

Pre-epoch: the v7 formula is fixed, but calibration remains provisional. Pilot results are not recommendation-eligible.