Bits against speed
Every nanocodec preset and every reference encoder, each in its own real-time rate control on one CPU core, compared with one zero line. Lower is fewer bits for the same quality; right is faster. Pick the clip set, the rate mode, the metric and the machine.
Our own measurements. The score is BD-rate on PSNR-YUV, VMAF v1 and MS-SSIM in equal thirds
(J); no subjective test was run. The held-out clips are videos none of our tuning used (most come from the same public collections as the
development clips); the development
clips are the ones we tune beside and flatter most of our presets. The references are mature, general-purpose
encoders measured only at their low-delay, one-thread settings; their other modes may do better. 8-bit 4:2:0
only. The Slow presets are on all three clip sets: both held-out sets measured on 2026-11-15 at build
bad14976; the development set at the builds the caption lists. How it is measured, versions and command lines.
This view as a table
More detail
By codecEach format's presets beside the references for that format.
SpeedEncode on Arm and x86, the browser build, hardware encoders.
Rate controlCoding efficiency at fixed QP against what real-time CBR costs.
Fixed QPEvery encoder at fixed quantizers on the AOM low-delay test conditions, the standards' way.
Spatial scalabilityAV1 and VP9 SVC against simulcast and libaom / libvpx.
MethodClips, settings, versions, command lines, window length.
YardsticksWhich independent implementation each speed claim is checked against.