Content-Adaptive Codec Selection Between AV1 and H.264: A Controlled Case Study of Bandwidth Savings Versus Latency at Matched Quality

Abstract

AV1 often reduces bitrate relative to H.264/AVC, but the gain is operationally useful only when the extra computation does not violate live-encoding constraints. This paper presents a controlled case study of content-adaptive codec selection using an openly available 640×360, 120-frame Big Buck Bunny excerpt. The publicly distributed source asset available to us is a GIF excerpt, which is decoded once to a Y4M reference; the study therefore characterizes transcode behavior on an openly distributed sample rather than first-generation compression from a production master. FFmpeg/libx264 (preset=veryfast) and FFmpeg/libsvtav1 (preset=10) are evaluated across CRF sweeps using bitrate, PSNR, SSIM, and wall-clock encode/decode time. All derived quantities in this revision were recomputed from the tabulated measurements. Over the overlapping SSIM interval 0.935–0.961, cubic-fit Bjøntegaard delta-rate (BD-rate) is −26.5% for AV1 relative to H.264, whereas over the much narrower PSNR overlap 32.91–34.20 dB the PSNR-based BD-rate is +59.7%; the latter should be interpreted cautiously because the overlap is narrow and the AV1 PSNR curve is comparatively flat. At a near-matched-SSIM point (H.264 CRF 20 versus AV1 CRF 32), AV1 lowers bitrate from 2286.1 to 1806.0 kbps (−21.0%) while increasing encode time from 1.299 to 4.444 s (3.42×) and decode time from 0.600 to 0.826 s (1.38×). Segment-level analysis over four 1 s chunks shows that AV1 is bitrate-efficient throughout, but one higher-motion segment falls below the adopted SSIM floor. Under a 2.0 s per-segment encode-time budget and SSIM ≥ 0.94, a simple policy selects AV1 for three segments and H.264 for one, reducing bitrate from 2416.4 to 2295.9 kbps while increasing mean SSIM from 0.948 to 0.954 relative to an all-H.264 baseline. The main conclusion is practical: for this content and configuration, AV1 is attractive when processing budgets are relaxed, whereas H.264 remains the safer choice when real-time feasibility is the primary constraint.

, , , ,