Skip to content

Tags: deeplabua/deepshrink

Tags

v0.3.15

Toggle v0.3.15's commit message
fix(core): a kept-as-is result never overwrites another file; release…

… 0.3.15

The kept-as-is copy lands under the output name with the source's extension
(book.shrink.mp3 for book.mp3), a name the CLI's collision check never saw:
an existing file there was replaced even without --overwrite. The CLI now
refuses it like any other collision, and the core creates the copy with
create_new so a library caller can't lose a file either.

v0.3.14

Toggle v0.3.14's commit message
feat(core): Discord preset 10 MB; slack and imessage presets; release…

… 0.3.14

Discord's free upload limit is 10 MB (the desktop catalog already used it);
the core still said 8 MB. Adds slack (1 GB) and imessage (100 MB, unpublished
by Apple - larger videos are sent as links) for the iOS Send-to sheet.

v0.3.13

Toggle v0.3.13's commit message
fix(core): hardware-encoder sample bias for size previews; release 0.…

…3.13

Quality-mode previews scaled sampled sizes by SAMPLE_BIAS (0.92), measured on
x264, whose keyframe-led samples over-predict. Apple's encoder barely does:
on 6 clips x H.264/HEVC the unbiased samples ran 0.88-1.12 of the real size
(mean 1.03), so 0.92 under-predicted by ~6 % on average (8/12 within +-10 %).
SAMPLE_BIAS_HW = 0.97 for hardware plans: 11/12 within +-10 %.

v0.3.12

Toggle v0.3.12's commit message
refactor(core): sample-size prediction in engine::plan; release 0.3.12

sample_secs / SAMPLE_BIAS move from engine::media to engine::plan, with the
window choice (sample_windows) and the size formula (predicted_bytes) split
out of predict_crf_bytes — pure, so the iOS encoder predicts like ffmpeg's
path. Behaviour unchanged; test for the windows and the formula.

v0.3.11

Toggle v0.3.11's commit message
chore(release): 0.3.11