mirror of
https://github.com/heygen-com/hyperframes.git
synced 2026-09-01 19:42:03 +00:00
The documented caveat was wrong. Probing which reference frame the live seek lands on shows it lands exactly on the right one (peak SSIM at frame 210 for t=7, falling off on both sides), so there is no seek drift to work around. The real floor is the decode gap: a replica is a live browser paint, a reference is a decoded compressed video. Flat graphics self-compare at 0.998-0.999; photographic video sits near 0.93 at high quality and 0.89 at draft, and most of that is a uniform level shift rather than a structural error. Adds meanSignedDiff so that shift is a number instead of an eyeball call on the overlay, printed as 'diff X% (bias +Y%)', and replaces the caveat with the measured per-content floors.