Optimizing x264 settings and per-title ladders
Jan Ozer tests x264 optimization and per-title encoding ladders, finding a 33% bandwidth reduction and 0.66 VMAF quality increase at the cost of 3.2x longer encoding times.
read it Detailed testing data and parameter-by-parameter analysis provide actionable benchmarks for video engineers optimizing H.264 pipelines.
summary
Video engineer Jan Ozer benchmarks x264 encoding settings and per-title ladders across thirteen test clips to maximize quality while minimizing bandwidth. By combining a ten-second GOP, two-pass encoding, three reference frames, and a per-title ladder targeting a VMAF score of 93, the configuration reduces average bandwidth from 4.9 Mbps to 3.27 Mbps while slightly improving VMAF. Although encoding time triples, bandwidth savings offset the extra compute costs after approximately 186 viewing hours. The analysis details the isolated impact of individual parameters like presets, GOP size, and reference frames on both compression efficiency and encoding speed.
key points
- Optimizing x264 settings and per-title ladders reduced average bandwidth by 33% (from 4.9 Mbps to 3.27 Mbps) while increasing VMAF by 0.66 points.
- Encoding time increased by 3.2x, but bandwidth cost savings covered the extra compute expense after 186 hours of viewing time.
- A ten-second GOP improved VMAF by 0.9 points over a two-second default, while reducing reference frames from 16 to 3 clawed back significant encoding speed with minimal quality loss.
who should read it
Video streaming engineers and backend architects managing H.264 encoding pipelines and bandwidth costs.
discussion mixed
consensus
Commenters discuss the practicalities of H.264 versus newer codecs, noting that legacy client support often forces the continued use of H.264.
pushback
Some commenters question VMAF as a reliable metric for perceived video quality, recommending alternatives like SSIMULACRA2 or CVVDP instead.
I thought VMAF was a bad metric for perceived video quality.
CVVDP is best and SSIMULACRA2 and Butteraugli pretty good.
It highlights the ongoing debate over whether VMAF is an accurate metric for perceived video quality compared to alternatives like SSIMULACRA2.
Generated by an LLM from the article and thread. Check the sources before relying on it.