Convex hull and per-title encoding
A streaming service encodes each title several times, at different sizes and bitrates, so a player can pick what its connection carries. That list is the bitrate ladder. A fixed ladder uses one table for every title. Per-title encoding measures each title and builds its own table, because the bitrate at which a larger picture starts to pay off depends on the content.
Find the best size for a bit budget
The cat is coded at three sizes: 32×32, and halved once and twice before coding. Each size is coded at 13 values of the quantization parameter (QP), and every decode is scaled back to 32×32 and compared with the full-size source by peak signal-to-noise ratio (PSNR). Bits are a modeled coefficient cost, not a written stream, and every number here, the switch points included, belongs to these 39 samples and that price list. Set a budget: the table gives the best encode of each size that fits, and the ring marks the best of all. The wide band is the best sample within each budget. It is a staircase: between two samples nothing else was coded. Its corners are the nondominated samples, those that no other sample matches or beats in quality with as few bits. Netflix's post on per-title encoding calls the boundary where each size is best the convex hull; a strict convex hull can leave out more samples, as the line under the plot counts.
| Size | QP | Bits | PSNR, dB |
|---|---|---|---|
| 32×32 | |||
| 16×16 | |||
| 8×8 |
For the cartoon cat at 600 bits:
- The best sample is 16×16, at 23.8 dB for 544 bits, and no 32×32 sample fits. The answer stays there for every budget up to the next corner of the staircase. Among these samples full size leads from 793 bits.
- A smaller size levels off. The cartoon cat at 16×16 levels off near 24.3 dB with this way of halving and scaling back: a finer quantizer cannot return what halving removed.
- The soft cat stays small for longer. Its 16×16 encodes reach 35.9 dB and 32×32 leads only from 1146 bits. At a budget of 1000 bits the cartoon cat is best at 32×32 (27.1 dB against 24.1) and the soft cat at 16×16 (35.5 against 34.1). One switch point cannot be right for both.
- The same modeled bits give different PSNR against each source. At 2000 bits: 37.0 dB for the cartoon cat, 43.6 for the soft one, 32.2 for the grainy one. Each is scored against its own source, the soft cat against the blurred picture, so the three are not one scale of how good a picture looks.
From curves to a ladder
Netflix described the procedure in 2015: trial encodes at several sizes and quantizers, and a ladder built from the points closest to the hull.
The efficient encodes are a starting point, not the ladder. A service still chooses how many rungs to offer, how far apart, and which sizes its devices can decode: the post aims for neighboring rungs a little under one just-noticeable difference apart and adds ladders for devices limited to smaller sizes. Such limits can call for a rung that is not among the efficient encodes.
What the Netflix post says, and how quality is measured at a smaller size
For each trial encode the post measures bitrate and quality, and it interpolates a curve for each size before picking the ladder's points. This demo interpolates nothing: its budget picks only encodes that were made. The post observes what the demo shows: each size's quality levels off as bits are added, a larger picture can score below a smaller one at the same bitrate, and different sizes are best over different ranges of bitrates; nothing guarantees every size a range.
For quality at a smaller size, the documentation of Netflix's Video Multi-Method Assessment Fusion metric says to scale the decode up to the reference's size and measure there, as this page does.
Trade-offs
- The trial encodes are the cost. This demo ran 39 encodes for one picture. The analysis is repeated for every title.
- The hull belongs to the metric and the scaler. Another metric or another way of scaling back up moves the crossovers; neither was varied here.
- Nondominated is not the same as on a convex hull. Which samples no other beats does not change with the scale of the axes. A strict hull also drops samples under the straight chords between its points, and that does change with the axes: for the cartoon cat on linear axes it keeps no 16×16 encode, though one is the best sample at 600 bits. A chord is not an encode you can pick.
- A title is not uniform. One curve per title averages easy and hard scenes; nothing here measures that.
Sources
Read 2026-11-15 and 2026-11-16.
- Netflix Technology Blog, Per-Title Encode Optimization, dated 14 December 2015. Read as extracted text through a fetch tool, and again on 2026-11-16 for how it describes the hull, the interpolated curves, the spacing of rungs and the ladders for devices limited to smaller sizes; its tables and figures did not come through, and the text does not say how quality was measured at smaller sizes.
- Netflix's VMAF repository at commit a6ed0f1: scaling a smaller decode up to the reference's size in
resource/doc/faq.md. - Not checked: Netflix's paper on the method and any production ladder; how a production ladder is chosen beyond what the post says; per-shot encoding; any encoder, metric or scaler other than the demo's; any viewing test.