Skip to content

v0.2.1 — lively narration, Gemini alignment and dialogue, 28 styles

Choose a tag to compare

View all tags
@ZLHad ZLHad released this 30 Sep 15:02
· 9 commits to afc0d375eeb54aba6909ca2d63f86aea061073e8 since this release
v0.2.1
7014293

Narration with feeling and rhythm (user: "声音是对的，但不够活泼，太僵硬")

  • bin/vh tts accepts a per-line direction in [brackets], added to the overall --instruct for that line only and kept out of captions.

  • --beats map.json [--snap beat|half|downbeat] [--lead s] starts every line on the next grid point instead of a fixed gap. @id:downbeat pins one line, e.g. the answer on the drop. Tested live with Gemini at 120 BPM: lines landed at 0.50 / 4.00 / 5.00 / 10.00 s exactly.

  • playbook/04 "让声音有表情、有节奏":

    • how to direct narration, with a default delivery per video type;
    • riding the beat;
    • frame-aligned tempos at 30 and 24 fps;
    • rhythm density per type;
    • mixing so music stays present (a measured duck_ratio=3 hole vs a clean 1.6);
    • writing sound as prompts.

    Types 02, 03, 04 and 08 point to it.

Gemini narration, further

  • Gemini 3.8 Flash TTS is now tested live in Mandarin and English (it was only mock-tested in v0.2.0). Also documented: the default local Qwen3-TTS 0.6B sometimes runs on after short English lines.
  • --align gemini: every line, whatever the provider, is transcribed by gemini-3.5-transcribe for word timestamps and compared with the script. A line below --min-sim (default 0.85) is flagged, which catches the Qwen 0.6B run-on and skipped words. --vocab passes terms to the transcriber.
  • Two-speaker dialogue: an @speakers A=Kore B=Puck header, then A: … lines. Labels only count when declared, so an ordinary "注意：" line stays narration.
  • --join block|all sends consecutive lines as one request, so the delivery flows across sentences; the per-line files are then cut at word timestamps. A Gemini dialogue defaults to block.
  • bin/vh voices list | design | delete manages Gemini designed voices.
  • bin/vh captions writes per-word timing into captions.json and a captions.<lang>.lines.srt with one cue per wrapped line, each starting on its first spoken word.
  • The key is read from GEMINI_API_KEY only; nothing is written to the repo.

Reading time and motion rules

  • tools/readcheck.py / bin/vh readcheck: on-screen text needs max(2.5 s, CJK ÷ 4.5 + other ÷ 15 + 1.5 s), adapted from lemo-opuscar (MIT); subtitles follow the Netflix ceilings (CJK ≤ 9 chars/s, English ≤ 20) with a 1.8 s floor. Checklist item #5 and playbook/03 use it.
  • playbook/03: four spring registers by element (buttons, cards and camera, big type and logos, mascots), from xilo-opus-video, plus the NaN trap at ζ = 1.
  • playbook/08: sub-frame motion blur must not sample across a cut, and per-subframe sub-pixel jitter gives free anti-aliasing (from abstract-algebra-promo).
  • playbook/02: CJK web fonts on canvas arrive as unicode-range subsets, so document.fonts.ready can pass with glyphs missing. Call document.fonts.load once per weight with every character the film uses, and fail loudly on rejection, an empty result or a timeout.
  • playbook/05: green-screen character + code scene: key and despill with ffmpeg, keep alpha (PNG or ProRes 4444), let code own the scene, type and beat.

Styles: 28

  • Two new presets, each with a swatch, QA and determinism check, and two rounds of independent review: pixel-16bit (Chrono Trigger, A Link to the Past; 320 × 180 integer scaling, a fixed 24-color palette, mosaic transitions; lowest score 7) and isotype (Neurath and Arntz; one symbol = a fixed quantity, added or removed on the beat; lowest score 6).
  • The deferred fixes from v0.2.0 are done: monumental-scifi has a hook, cutout-jazz and swiss-grid-type no longer share a motif, and the shadow-puppet click is gone. gallery.jpg / gallery.mp4 rebuilt for 28.
  • lemo-opuscar is now 43 styles and MIT for the whole repo (since 2026-09-29; older snapshots were CC BY 4.0). Presets adapted from the older snapshot keep their CC BY attribution.

References

  • cases/opus55-gallery.md §7: a deep read of abstract-algebra-promo (sub-frame blur, cut protection, math promo structure); §5 adds the dsxzai catalog.
  • references/fetch.sh: text-only clones now actually pull on update (the empty checkout used to short-circuit it).

Full history: CHANGELOG.md

Assets 2
Loading