Skip to content

v0.6.0「看得见画面、接得上生态」/ See the picture, join the ecosystem

Latest

Choose a tag to compare

@github-actions github-actions released this 10 Jul 08:53

v0.6.0「看得见画面、接得上生态」/ "See the picture, join the ecosystem"

11 项新能力,比 v0.5 更厚的一版:AI 终于"看得见画面",切完的片子直接接上你的发布与精修生态。

🧠 六路证据链——爆点判断不再只靠文字

  • 视觉爆点信号:可选接本机 Ollama 视觉模型(如 qwen3-vl:4b)抽帧研判画面高能时刻——夸张表情、激烈动作、场面炸裂这些逐字稿里看不见的爆点;抽帧押注响度/镜头信号窗口,端点不可用自动跳过
  • 表情峰值信号:YuNet 人脸 + FER+ 表情识别(两个几 MB 的 MIT 模型,首次自动下载),零配置圈出大笑/惊讶/激动时段——不装 Ollama 也能让 AI"看见"情绪
  • 弹幕热度信号:自动发现录播旁同名弹幕 .xml(录播姬约定),滑窗密度 + 高能词加权圈出观众实时高能时刻——弹幕是观众逐秒投的票,证据力最强;阈值随全场基线自适应
  • 与既有的 文本 + 响度峰值 + 镜头密度 组成六路证据,全部 fail-open

🌍 双语字幕 + 出海交付

  • 双语字幕烧录:LLM 整句翻译(整句语境,不是字幕行碎片)烧成主字幕下方小号译文轨;中文源译英、外语源译中;跳剪时间轴同步重映射
  • SRT 字幕文件:每条切片旁同名 .srt,时间轴对齐跳剪成片、断行与烧录字幕逐行一致,双语含译文行
  • 发布文案:每条切片生成钩子标题 + 垂类话题标签 + 简介,落 .post.txt 直接复制粘贴

🤖 接上生态

  • 本地 MCP Server:在 Claude Code / Claude Desktop 注册后,一句"把这个录播切成爆点"驱动整条本地管线(clip_video / detect_highlights / transcribe_video),素材不出电脑
  • 录播监听:盯住录播姬/OBS 输出目录,新录播写完落稳自动转写、找爆点、出成片——7×24 无人值守,写盘中不切半截,重启不重切
  • 时间线 EDL 导出:AI 切点(含跳剪每一刀)写成 CMX3600 时间线,DaVinci/Premiere 重链源片继续精修——AI 粗剪,最后一刀永远是你的

✨ 顺手的小事

  • 切片时长档:短·10-30s / 标准 / 长·40-90s 一键切换重检测,目标时长硬约束进选段提示词
  • 出片偏好记忆:开关组合自动记住,换视频不再重新点一遍
  • 标题即点即改:候选标题原地编辑,文件名/贴片/发布文案全部跟随

底线不变

本地处理、素材不出电脑;真免费、无水印、不限时长;每个 AI 决定可见、可改、可推翻。


11 new capabilities — a bigger release than v0.5: the AI finally sees the picture, and finished clips plug straight into your publishing & editing ecosystem.

🧠 Six-signal evidence chain

  • On-device vision peaks (optional Ollama VL model samples frames for visually hot moments), facial-emotion peaks (YuNet + FER+, zero-config, tiny MIT models) and live-chat density (auto-discovered danmaku XML, adaptive threshold) join text + loudness + shot density — all fail-open

🌍 Bilingual & go-global kit

  • Bilingual caption burn-in (sentence-level LLM translation as a second track, jump-cut-aware), SRT export (burn-in-consistent, bilingual-aware) and post copy generation (hooky title + niche hashtags + description per clip, saved as .post.txt)

🤖 Ecosystem-native

  • Local MCP server — one sentence to Claude drives the whole local pipeline; 24/7 recording watch folder — recordings auto-clip the moment they finish writing; EDL timeline export — every AI cut point (each jump-cut splice included) lands in DaVinci/Premiere for the human final cut

✨ Quality of life

Clip-length presets (short/standard/long as a hard prompt constraint), render-prefs memory, inline title editing.

Unchanged

Local-first, footage never leaves your machine; genuinely free, no watermark, no length caps; every AI decision visible, editable, overrulable.

⚠️ 安装包未签名:macOS 首次打开需右键→打开(或 xattr -cr /Applications/HotClip.app);Windows SmartScreen 点"仍要运行"。/ Unsigned builds: macOS right-click → Open (or xattr -cr); Windows SmartScreen → "Run anyway".