Per-Title Encoding: Quality, Bitrate & Ladder Design | Callaba

Per-title encoding builds a bitrate ladder from the measured behaviour of one video instead of forcing every asset through the same resolution and bitrate table. The practical goal is not “maximum quality.” It is to stop spending bits where viewers cannot see a useful gain while protecting scenes that fail under a generic ladder.

One ladder cannot describe every title

Animation with clean edges, a grainy concert, a screen recording and a fast sports sequence stress an encoder in different ways. At the same bitrate, they can land at very different perceptual quality. At the same quality target, they may need different bitrates or even different useful resolutions.

Start with a rate-quality experiment

The original Netflix per-title description derives a title-specific ladder from rate-quality curves for multiple resolutions. That framing is more useful than copying someone else's final bitrate table: encode comparable samples across candidate resolutions and rates, measure each output, then identify points that are not dominated by a lower-bitrate or higher-quality alternative.

Compare like with like: use identical source intervals for every candidate. Include opening credits, faces, low-light texture, fast motion, gradients, grain and any scene known to challenge the codec. If the hardest sequence appears near the end, a quiet first minute tells you very little.

Use metrics as evidence, not as the viewer

VMAF is a full-reference perceptual quality metric with published models and tooling. It can rank candidate encodes efficiently, but the score depends on the model, scaling path, reference preparation and viewing assumptions. Record all of those inputs with the result. Do not compare scores produced by different hidden preprocessing chains as if they share one scale.

Add targeted visual review around the decisions that change the ladder. Inspect dark scenes, fine text, flashes, scene cuts and texture that a pooled score can conceal. Confirm audio, captions and A/V synchronization as separate acceptance checks; a strong video metric says nothing about a missing language track.

Build the frontier before choosing the rungs

Plot measured quality against actual delivered bitrate for every candidate. Remove a point when another candidate provides equal or better measured quality with fewer bits, or equal or lower bitrate with better quality. The remaining points describe the useful rate-quality frontier for this title.

Only then choose the rungs, with real playback in mind. Leave enough separation for the adaptive player to make a meaningful switch. Near-duplicates add storage, encoding and manifest complexity without giving the player a useful choice. Preserve required device limits, codec profiles and maximum dimensions even when a laboratory result suggests a more efficient but unsupported candidate.

Set quality bounds and operating bounds together

A quality threshold alone can produce an impractical ladder. Add limits for maximum bitrate, rendition count, resolution, decode level, segment alignment, encode time and storage. For a paid archive, a small visible improvement may justify another rung. For frequently replaced clips, the same gain may not repay its compute and operational cost.

Define the low end deliberately. It should remain decodable and understandable on constrained connections, not simply be the last candidate that survived a metric threshold. Check faces, text and essential action at the actual display sizes expected for that rung.

Measure savings across the catalogue

Per-title encoding moves work from a simple preset into analysis, candidate encoding, metric computation and approval. Report total delivered bytes, encoding minutes, storage, rejected candidates, quality distribution and playback errors by title class. Savings on easy animation can subsidize higher rates for difficult material, but only if the pipeline avoids producing every candidate forever.

Cache source fingerprints and analysis results. Re-run when the source, encoder, metric model, codec policy or target device matrix changes. A ladder selected with one encoder version is evidence for that version, not a permanent property of the movie.

Release a ladder with reproducible evidence

1. Fingerprint the input. Preserve the source hash, duration, frame rate, colour and audio properties.
2. Version the experiment. Record encoder build, arguments, metric model and preprocessing.
3. Review the frontier. Explain each kept rung and every policy override.
4. Validate packaging. Check aligned boundaries, manifests, tracks and representative playback.
5. Watch production. Compare startup, rebuffering, switches, errors and delivered bytes with the previous ladder.