kajisho5/ffmpeg-skill v3.8.12

on GitHub

Automated release: version and notes generated from pull requests merged since 3.8.11.

Fixed

• A window or a length on the picture keeps the frame it names (#470's open items).
  overlay.py --start/--end, graphics.py's windows, broll.py's cutaway and fit.py --method trim without sound wrote a frame's time to the millisecond, which at 30000/1001 is past the
  frame (frame 37 at 1.2345667 s is "1.235"). broll.py --at 1.2345666 --from 1.2345666 --duration 1.0343666 put B's frames 39-68 on frames 38-67 and puts 37-67 on 37-67; an
  overlay.py window of frames 37-67 is on 37-67, where a 25 fps --image/--video layer was
  on 36, 38-65 and 67 (overlay's enable is evaluated on the layer's frames too, so the window
  is in the layer's x position now) and --text on 38-67; fit.py --method trim --duration 1.2345666 writes the 37 frames asked, not 38. The bounds sit half a frame from the frames
  they select, so these commands change for windows in whole milliseconds too. An SMPTE time at
  23.976, 29.97, 47.952, 59.94 or 119.88 is counted at the exact N * 1000 / 1001 (00:00:01:07 was
  1.2345679 s, past frame 37). Not changed: the emoji PNGs caption.py and graphics.py's ASS
  route lay over a libass line, and fit.py --method speed without sound (38 frames for 37 on
  FFmpeg 6.1.1).
• A cutaway's sound, broll.py --audio b's mute and sync.py --trim-second's delay are
  placed to the sample. --audio b muted the A-roll
  per 1024-sample audio frame (48128-96255 for 48000-95999) and is evaluated per sample now. A
  cutaway's sound starts with its first frame, the A-roll's first at or after --at (--at 1.001
  at 25 fps: 1.040 s, where the sound started at 1.001 s and the picture at 1.04 s). sync.py --trim-second --fine-ms 0.25 started the second file's sound at sample 11232 for an offset of
  0.2343 s and starts it at 11246; at the default --fine-ms 1 without --fix-drift the command
  is unchanged. fit.py --method speed's retimed sound keeps the length atempo gives it (some
  milliseconds off the target): padding it to the sample did not end on FFmpeg 8.0.1.
• pad.py --start and freeze.py --at 0 --mode insert delay the sound by the frames put ahead
  of the picture. The picture is padded by whole frames: --start 0.23425 at 25 fps is 6
  frames (0.240 s), and the sound was delayed by 0.23425 s, 5.75 ms ahead of its picture; it is
  delayed by 0.240 s now (--start 0.05 at 30 fps: 0.0667 s for 0.050 s). The silence after a
  clip is unchanged.
• loop.py fills the gap a sound that outlasts the picture leaves before each repeat on the
  clip's own rate. On FFmpeg 6.1.1 the next repeat's first frame was shown twice, from a frame
  before its sound; the graph now repeats the last frame and puts each repeat's frames on the
  grid once, so the frames written no longer depend on FFmpeg's default frame-rate handling
  (7.1 and 8 held the last frame instead; not measured locally). A clip without that gap writes
  the command it did.

All of it measured on FFmpeg 6.1.1 only (docs/design-decisions.md, "After 3.8.11").

• fix: a window, a start or a length lands on the frame or the sample it names (#514)

What's Changed

• fix: a window, a start or a length lands on the frame or the sample it names (#470) by @kajisho5 in #514

Full Changelog: v3.8.11...v3.8.12