Reddit reviews Video Encoding by the Numbers: Eliminate the Guesswork from your Streaming Video

There is a reason that the presets are called ultrafast-faster-fast-medium-slow-slower rather than ultrabig-bigger-big-normal-small-smallest. These are the x264 project's presets focused on encoding speed/overhead and H.264 complexity, rather than bitrate (or even quality). CRF is about quality. VBV bitrate is about size and bandwidth.

Constrained CRF with a VBV cap (technically called the HRD model in the H.264 specification) is an excellent way to maintain quality while controlling size or streaming bitrate. I find that setting a bitrate calculated across a 6s buffer is an good starting point to control bitrate or filesize.

-crf "23" -maxrate "2000000" -bufsize "(6.000*2000000)/8"

bufsize is in bytes, maxrate is in bits. Hence the 6.000s and 8 bits in a byte.

This method has never let me down when encoding for HLS / DASH / CMAF for streaming as well as MP4 file-based preparation.

If livestreaming then probably replace -crf with -bitrate:v, with 10% breathing space.

-bitrate:v "2000000" -maxrate "2200000" -bufsize "(6.000*2000000)/8"

(Although OZER, Jan discusses and Apple allows for up to 200% for live streaming.)

The x264 people (kudos Team x264 and Team FFmpeg) and a bunch of cube-farm drones in Cupertino have put hours into coming up with pretty solid guidelines which meet generalized use-cases. The Scene folks take optimization to another level.

I recommend learning to use VMAF, PSNR & SSIM to measure objective quality. As well as reading https://www.amazon.com/Video-Encoding-Numbers-Eliminate-Guesswork/dp/0998453005.