Skip to content

Research-0068: vmaf-tune Phase E — per-title bitrate-ladder algorithm survey

  • Date: 2026-05-03
  • Author: Lusoris (with Claude assistance)
  • Companion: ADR-0295

Question

Given a single source clip and a sweep of (resolution, target-VMAF) samples, what is the right algorithm to pick n ABR ladder rungs?

Survey

Netflix per-title encoding (2015 paper)

Sample (resolution, bitrate) on a grid; compute the upper convex hull of (bitrate, vmaf); the hull is the ladder. Industry baseline. Two properties that matter:

  1. Pareto frontier — for any bitrate, the hull rendition has the maximum VMAF; nothing else dominates.
  2. Diminishing returns — the slope (Δvmaf / Δbitrate) is monotonically non-increasing along the hull. ABR clients stepping up rungs see ever-smaller quality gains for ever-larger bitrate spends.

Apple HLS Authoring Specification §2.3

A fixed recommended ladder, geometric in bandwidth (×2 per rung). Trivial to implement; useful as a default for content where the source isn't known. Equivalent to ignoring the title — defeats the per-title premise.

JND-spaced (Visicom 2019, JND-VMAF / JND-aware ABR)

Pick rungs so each adjacent pair differs by ~6 VMAF (one just-noticeable-difference). Perceptually motivated; requires a JND model. We don't have one — would layer on top of VMAF if we did. Deferred.

Bayesian-optimisation sampler

Replace the resolution-grid × target-vmaf sweep with sequential BO querying. Fewer encodes per title; principled exploration. Orthogonal to the picker — the ladder math runs on whatever sample cloud the sampler produces. Deferred.

av1an --target-quality

Per-rendition bisect. Conceptually a Phase B sibling, not a Phase E sibling — the sampler half of the pipeline. We already wire that as Phase B (PR #347).

Decision

Implement the Netflix two-pass approach: Pareto filter, then upper- convex envelope with cross >= 0 pop predicate (drops accelerating- returns interior points so the hull is everywhere concave / diminishing-returns). Pluggable sampler — production wires Phase B's bisect; tests inject a synthetic stub. Default rendition picker uses log-bitrate spacing (Apple HLS authoring spec convention) with VMAF spacing as an opt-in for perceptual ABR work.

Open questions

  • Real-corpus PLCC of the picked ladder against Netflix's published per-title rungs. Gated on Phase B merging (so we have a real sampler) plus a Netflix-Public encode of N representative titles. Deferred to a follow-up validation digest.
  • Whether to add a JND-spaced rung-picker once we ship a JND head (tiny-AI roadmap). Out of scope for Phase E.

References

  • Netflix per-title encoding paper (2015) — the canonical reference.
  • Apple HLS Authoring Specification for Apple Devices §2.3.
  • av1an / ab-av1 — prior art for --target-quality per-rendition bisect (Phase B, not Phase E).
  • Bitmovin Per-Title — closed-source equivalent.
  • PR #354 capability audit (docs/research/0061-vmaf-tune-capability-audit.md, pending) Bucket #6 — the source flagging this as the fork's highest-leverage gap.