Skip to content

v0.12.0「挑得准、开场炸」/ Pick sharper, open louder

Choose a tag to compare

View all tags
@github-actions github-actions released this 05 Aug 11:28
· 45 commits to main since this release
v0.12.0
2c93f20

v0.12.0「挑得准、开场炸」/ "Pick sharper, open louder"

这版把力气花在「挑哪段」和「头三秒」:AI 挑段第一次有了「眼睛」复核画面,开场多了爆点闪现这张牌,切点精度补上快档转写的短板;字幕默认样式也按 2026 主流审美换了代。

⚡ 爆点闪现(flash-forward)

正片开始前,先闪 0.3-1 秒全片最炸的瞬间,再切回从头讲——「等会儿有好东西」的悬念开场。

  • 全网仅 0.04% 的切片做了 visual hook,这是条几乎没人走的差异化路
  • 与高潮前置(cold-open)自动二选一,不会同片叠两种开场
  • 峰值离片头太近、或找不到安全切点时自动跳过——宁可不做不可做错

👁️ AI 画面复核

检测完成后,分数靠前的候选各抽一张九宫格接触表,让视觉模型过目再排一次序:

  • 画面炸裂加分:文字稿平平但画面起飞的段子(整活/名场面)不再被埋没
  • 信号候选画面死气降分:纯靠声音信号顶上来、画面却是静态空镜的候选被拉下来
  • 货不对板标警告:画面与标题对不上的候选,推荐理由里明确提醒
  • 每条候选只花一次 VLM 调用;本机 Ollama(qwen3-vl:4b)免费,视觉设置里填个 API Key 即切云端模型,零新增配置
  • 全程 fail-open:视觉模型不可用时检测结果原样返回

🎯 精准切点

导出前用 Paraformer 对选中片段做二遍解码,拿它的一体化词级时间戳(±50ms 级,精度超过强制对齐)修正字幕、跳剪与切点:

  • 快档转写(SenseVoice)听得准但时间戳粗,这个短板由二遍对齐补齐——两个引擎各干各的强项
  • 主转写已是 Paraformer 档时自动跳过;首次使用自动下载 ~240MB 模型
  • 对不上(如背景音乐段、对齐置信度不足)自动回退原时间戳,绝不改坏

🪝 钩子兑付校验

钩子里承诺的数字(「三个方法」「省 200 块」)必须真出现在片中——没兑付的写进质检报告,防止 AI 无意间产出「标题党切片」。

✨ 字幕换代

  • 默认样式换「关键词高亮」:一屏一短句 + 关键词品牌色,中文短视频的主流形态
  • 老「逐字点亮」(整句扫色卡拉OK)已过时,降为备选样式继续保留
  • 「逐词弹色」升级:当前词品牌色点亮、说到哪亮到哪(提前 80ms 跟手),弹入改为阻尼手感不再夸张过冲——对齐当下逐词字幕的主流工艺

🔧 修复

  • 桌面端 4 个从未生效的导出开关接回:降噪/高潮前置/精华合集/一键双画幅(此前 UI 能点但导出层收不到)

⬆️ 升级方式

到 Releases 下载新安装包覆盖安装即可。模型、词表、偏好原位保留;已保存的字幕样式偏好不受默认值变更影响。

This release is about what gets picked and the first three seconds: candidate selection gains eyes for the first time, openings gain the flash-forward card, cut-point precision covers the fast ASR tier's weakness — and the default caption style catches up with 2026.

⚡ Flash-forward cold opens

Before the story starts, flash the clip's most explosive 0.3-1s, then cut back — the "something good is coming" tease. Only 0.04% of clips ship any visual hook. Auto-coordinates with the climax-first cold open (one or the other, never both), and skips itself when the peak sits too close to the start or no safe cut exists.

👁️ AI visual review

After detection, top candidates each get a contact-sheet look-over by a vision model: striking visuals boost the score, lifeless visuals demote signal-driven candidates, and title/visual mismatches get flagged in the reasoning. One VLM call per candidate; free with local Ollama (qwen3-vl:4b), or drop an API key into vision settings to use a cloud model — zero new configuration surface. Fully fail-open.

🎯 Precision cut points

Selected clips are re-decoded with Paraformer before export; its integrated word timestamps (±50ms, better than forced alignment) fix captions, jump cuts and cut boundaries. The fast tier (SenseVoice) hears well but times coarsely — this second pass covers exactly that. Skipped when the main transcript is already Paraformer; ~240MB model downloaded on first use; falls back safely when alignment is unreliable (e.g. music).

🪝 Hook payoff check

Numbers promised by the hook ("3 methods", "save $200") must actually appear in the clip — unfulfilled promises land in the QA report, so the AI never accidentally ships clickbait.

✨ Caption modernization

The default style is now keyword highlight (one short line per screen + brand-colored keywords — the current mainstream). The old full-line karaoke sweep is dated and demoted to an option. Word pop upgraded: the current word lights up in the brand color as it's spoken (80ms lead for that "on-beat" feel), with a damped entrance instead of the old exaggerated overshoot.

🔧 Fixes

  • Four desktop export toggles that were never wired up now work: denoise / cold open / compilation / dual-aspect (the UI accepted clicks but the export layer never received them)

⬆️ Upgrading

Grab the new installer from Releases and install over the old version. Models, hotword lists and preferences stay in place; saved caption-style preferences are unaffected by the new default.

Assets 9
Loading