Video Encoder Presets

Video encoder presets choose a bundle of coding decisions that trades encoding effort against compression efficiency.

An encoder can describe the same picture region in several valid ways. It must decide how much time to spend searching for a useful representation. An encoder preset packages those search and coding-effort choices.

Search another candidate

Use a toy decision cost J =D +0.1R, where D is a declared distortion measure and R is the number of coded bits.

Candidate Distortion D Bits R Cost J
A, found early 16 100 26
B, found after more search 4 120 16

Under this cost, B is better even though it uses more bits. Its lower distortion outweighs that cost. If the search stops after A, the encoder never gets the chance to select B.

Original arithmetic illustration of a decision objective, not an encoder benchmark, preset score or exact x265 configuration.

A different candidate could improve both size and reconstruction. More search creates opportunities to find a better tradeoff; it does not define “better” as the smallest possible file regardless of picture damage.

Which decisions take time?

Motion search evaluates possible reference positions. Partition search evaluates ways to divide a region. Transform and quantization decisions assess alternative corrections. Lookahead examines later pictures to plan relationships and resource allocation.

Faster presets take shortcuts or limit these choices. Slower presets can evaluate more candidates or use more detailed decision procedures. The resulting stream contains the choices selected, rather than the entire search history.

This explains why encoder computation and decoder computation are different quantities. A costly search can produce a relatively straightforward predictor. Some enabled tools also change decoder work, so the final tool/profile structure remains part of the output description.

What x265 medium and slow change

The published x265 table gives these selected examples:

Setting medium slow What the difference affects
Lookahead pictures 20 25 Amount of later picture analysis
Reference setting 3 4 Candidate retained views for prediction

These are selected settings from that documented implementation, not all the changes or a portable comparison with another encoder. Its presets also vary motion/partition and decision options.

The rate or quality target supplies the allocation objective. The preset supplies the effort used to serve it. A tuning changes priorities for a source/use case, such as grain or low latency. Keep these controls distinct when evaluating an encode.

Supply pictures fast enough for playback

Suppose an encoder can sustainably produce 20 output pictures per wall-clock second while the requested playback needs 30 per media second. Encoding one second of media takes 30/20 =1.5 seconds.

Under this model, the playback reservoir shrinks by 30 −20 =10 pictures per wall-clock second. A 300-picture prefill lasts 30 seconds of that continuing deficit. A short finite clip can finish within an adequate prefill; a continuous session cannot rely on a finite reservoir indefinitely. This is a declared throughput example, not the speed of any named preset or hardware.

Processing can have multiple frames in flight. The relevant requirement is sustained output throughput, rather than insisting that each individual frame's processing latency be shorter than 1/30 second.

For an Emby transcode, input decoding and required filters also share the processing job. A preset's useful additional search has to fit that complete supply path. An offline derivative can instead spend more time before the file is played, making a different effort tradeoff practical.

See Also

References