v0.5.0 —「选得准、批量稳」审阅台 · 镜头吸附 · 品牌模板 · 两级漏斗 / Pick right, ship steady

v0.5「选得准、批量稳」/ "Pick right, ship steady"

这一版全部围绕一件事:让 AI 选出的片,敢直接发。

This release is about one thing: trusting the clips the AI picks — enough to hit publish.

---

🎬 切片审阅台 / Clip review workbench

导出前先看片。每条候选一键打开审阅台——应用内直接播放这条切片(本地流式协议,拖进度条不卡),波形时间轴上拖两端手柄逐词微调切点(自动吸附字词边界,拖动即蹭帧),整句伸缩、一键还原 AI 切点,「查结尾」只播最后 2.5 秒验收收尾。你手调过的切点,导出时机器绝不再改。

Watch before you export. One click opens the workbench: play the clip right inside the app (local streaming, scrub-friendly), drag handles on a waveform timeline to fine-tune cut points word by word (snap to word boundaries, video scrubs as you drag), sentence-step extend/shrink, one-click restore, and a "check ending" button for the last 2.5 seconds. Manually tuned cut points are never overridden on export.

✂️ 镜头切点吸附 / Shot-snapped cut points

TransNetV2 镜头边界检测(31MB ONNX,本地推理,MIT)逐帧找真实镜头切换,切片起止点自动吸附到最近的镜头边界——成片不再从半个动作/半次转场开始。词边界守卫保证绝不切掉说话;检测失败自动回退;吸了多少写进 clips.json 回执。

TransNetV2 shot-boundary detection (31MB ONNX, local, MIT) finds real shot changes frame by frame; clip boundaries snap to the nearest one — no more clips opening mid-action. A word-boundary guard guarantees speech is never clipped; failures fall back silently; every snap is receipted in clips.json.

🎨 品牌样式模板 / Brand style templates

竞品锁在付费墙后的功能:主高亮色(卡拉OK点亮/关键词强调/开场钩子/气泡渐变同源)、字幕字号与位置三档、logo 水印(四角可选/透明度可调)。配一次存成命名预设,之后每条切片——包括一键托管——自动带上你的风格。

The feature competitors paywall: one highlight color driving karaoke sweep / keyword emphasis / opening hook / bubble gradients, three caption sizes and positions, and a logo watermark (any corner, adjustable opacity). Save named presets once; every clip — hands-off mode included — ships in your style.

🪣 端侧两级漏斗 / On-device two-stage funnel

可选用本机 Ollama 小模型(如 qwen3:4b)先通读全文圈出候选段,云端大模型只精读入围部分——长视频的云端 LLM 花费降一个量级。引用反查仍在全量转写上做,切点精度零损失;本地端点不可用自动回退全文,绝不影响结果;省了多少结果页明示。

Optionally let a small local Ollama model (e.g. qwen3:4b) shortlist candidate ranges first, so the cloud model only reads the shortlist — an order of magnitude less cloud LLM spend on long videos. Reverse matching still runs on the full transcript (zero accuracy loss); unreachable endpoints fall back silently; savings are shown in the results view.

---

一贯的底线 / As always

• 🔒 素材不出你的电脑(转写/镜头检测/人脸跟随/说话人分离全部本地)/ Your footage never leaves your machine
• 🧾 AI 做了什么全部写进 clips.json 回执,可审计 / Everything the AI did is receipted in clips.json
• 🆓 无积分制、无水印、不限时长 / No credits, no forced watermark, no length caps

安装包 / Installers: macOS (Apple Silicon dmg) · Windows (installer + portable zip)

⚠️ 安装包未签名(开源项目买证书太贵),macOS 首次打开需右键→打开;Windows SmartScreen 点「仍要运行」。/ Builds are unsigned; on macOS right-click → Open on first launch, on Windows click "Run anyway".

v0.4.3 — 工程底座:转写缓存 + 处理产出回执

可靠性与透明度 / Reliability & transparency

两块「完全托管」的地基:让重复使用更快,让 AI 的处理过程可审计。

转写结果本地缓存 / Persistent transcript cache

转写是整条流水线最慢的一步。结果现在按(文件 size+mtime,引擎)缓存在本地——同一个文件下次再打开、想换设置多切几条,跳过重转写、秒进挑爆点;云端引擎(ElevenLabs)命中缓存还顺带省一次 API 费用。文件被改动(大小/修改时间变化)或更换引擎会自动失效;缓存带 LRU 上限、故障静默回退绝不阻断真实转写。

Transcription is the slowest step. Results are cached locally by (file size+mtime, engine): re-open the same file to cut more clips and it skips straight to highlight selection; a cloud engine (ElevenLabs) even saves an API charge on a hit. Editing the file or switching engines invalidates automatically; the cache is LRU-bounded and fail-open.

clips.json 处理产出回执 / Processing receipt in clips.json

每条切片的 clips.json 现在多一个 render 字段,如实记录 AI 到底做了什么:有效字幕样式、取景是跟脸(face-track)还是回退到中心裁、跳剪的剪除比例与拼接段数、剪掉了几个口头禅。矩阵/CMS 流水线可据此审计,也能一眼看出哪条片子触发了降级回退。

Every clip's clips.json now carries a render block recording exactly what the pipeline did: effective caption style, whether the reframe followed a face or fell back to center-crop, the jump-cut ratio and splice count, and how many fillers were removed — auditable by CMS/matrix pipelines, and a quick signal for which clips hit a fallback.

---

本地处理 · 免 Key · 无水印。All local · no key · no watermark.

v0.4.2 — 字幕可读性:语义断行 + 防闪烁

字幕质量 / Caption quality

短视频字幕的「最后一公里」打磨——两处让自动字幕更像人工精修、更耐看。

语义断行(免 Key)/ Semantic line-breaking (keyless)

字幕不再按固定字数「拦腰截断」。顺着本地 ASR 标点模型标出的逗号/顿号在真实子句处换行;长句里没有标点时,再回看到最近的结构助词(的/了/着/过/地/得…)断,让「十几块的 / 到底有什么区别」这样的短语保持完整,而不是切成「…的到 / 底…」。等价于头部项目用 LLM 插 [br] 的效果,但用的是本地已有的标点信号——零额外模型调用、不需要云端 Key、每条片子都生效。

Captions now break at real clause boundaries: at ASR-detected commas, and — for long comma-free clauses — backed up to the nearest Chinese structural particle so phrases stay intact. Same intent as an LLM-inserted [br], but from the local punctuation signal: no extra model call, no cloud key.

防闪烁 / Anti-flicker

默认的卡拉OK/关键词字幕此前在每次行切换时会闪一帧空白(上一行在最后一个字结束时立刻消失,下一行还没到)。现在每行会保持到下一行开始再消失,连续语音之间不再有空帧闪烁;超过 0.8s 的真实停顿仍会正常清屏。气泡特效字幕也统一了同一行为。

The default karaoke / keyword captions flashed a blank frame on every line change. Each line now holds until the next one begins, so continuous speech never flickers; a genuine pause (>0.8s) still clears the caption. The bubble overlay was unified to the same behavior.

---

本地处理 · 免 Key · 无水印。All local · no key · no watermark.

v0.4.1 — 多人对谈字幕按人上色·端到端打通

修复 / Fixes

多人对谈字幕「按人上色」现已真正端到端 / Multi-speaker caption coloring now works end-to-end

v0.4.0 引入了说话人分离与气泡字幕的按人上色,但打了说话人标签的转写稿只用于 LLM 挑段、并未回传到导出环节,且导出时逐词映射丢弃了 speaker 字段——README 声称的「气泡字幕按人上色」在整链上并未真正生效。本版把这条链补齐:

• 检测接口回传打标签的转写稿(diarization-labeled transcript)
• 渲染端把它提升为应用状态,导出时随之携带
• sliceWords / 逐词映射保留 speaker 字段,气泡模板按 speaker % 调色板 上色(关键词渐变仍优先)

The speaker-labeled transcript is now returned from detection, lifted into app state, carried into export, and the per-word speaker id is preserved through sliceWords and the overlay payload — so bubble captions genuinely tint by talker in interviews / podcasts / co-streams.

顺带修一处潜在漏送 / Also fixed a latent gap

当同时关闭字幕与跳剪、仅保留「剪口头禅」时,转写稿不会随导出请求发出,剪口头禅会被静默跳过。已把 cleanFillers 纳入送稿条件。
Fixed: with captions and jump-cut both off but filler-cleanup on, the transcript was not shipped to export, silently disabling filler removal.

---

本地优先 · 零上传 · 无水印。All local, no upload, no watermark.

v0.4.0

v0.4.0 — 气泡特效字幕 + 四维爆款分 / Bubble FX Captions + 4-Dimension Virality Score

字幕进入 Web 渲染时代 / Captions enter the web-rendering era

• 气泡特效字幕:全新 Web 渲染字幕引擎——用应用内置的 Chromium 离屏逐帧渲染 CSS 字幕层,再合成进成片。自适应圆角气泡底、关键词渐变金字、弹性弹跳入场,这些传统 libass 烧录做不出的效果,现在导出页一档切换即得(逐字点亮/划重点/大字弹跳/气泡特效/无)。确定性逐帧驱动:同输入必出同片。
  Bubble FX captions: a new web-rendered caption engine — the app's own Chromium rasterizes a CSS caption layer offscreen frame-by-frame, composited into the final clip. Adaptive rounded bubbles, gradient keyword text, springy pop-in — effects libass structurally cannot produce, now one style-toggle away.
• 四维爆款分:AI 复评升级为 钩子/结构/价值/热点 四维分项打分,每维一句话理由;总分按批内排名归一(推荐档 76-99),同批可直接比大小,不受 AI 打分批间漂移影响。每条候选另附 ≤15 字悬念句(teaser),已进 clips.json 元数据。
  4-dimension virality score: hook/flow/value/trend sub-scores with one-line reasons each; final score is rank-normalized (76-99) so numbers are comparable within a batch. Each clip also gets a spoiler-free teaser line, exported in clips.json.
• 跳剪声学双重判定:剪不剪由「无词 且 声学静默」共同决定——笑声、掌声、BGM 高潮这些没有台词但有情绪的瞬间不再被误删。
  Jump-cut acoustic AND gate: a gap is only cut when it is wordless AND acoustically silent — laughter, applause and BGM stings survive.
• 云端转写第四档 ElevenLabs Scribe:90+ 语种、词级时间戳;只上传提取的音轨(不传视频),Key 存本机。
  ElevenLabs Scribe cloud ASR tier: 90+ languages, word timestamps; only the extracted audio track is uploaded, never the video.

完整技术细节见 README。/ See README for details.

v0.3.0 — 人脸跟随智能取景 / Face-tracking smart reframe

这一版的主角:人脸跟随智能取景。 竖屏 9:16 重构不再是傻居中——人在哪,画面跟到哪。

✨ 新能力

🎯 人脸跟随智能取景

• 竖屏重构自动检测人脸并跟随取景:人物走动、站位偏侧、多镜头切换,主体始终稳在画面里
• 镜头级三模式,「能不动就不动」:静止镜头稳如三脚架、缓慢移动平滑横移、真实走动用 One Euro 滤波跟踪(零抖动)
• 与去录屏UI、气口跳剪、字幕烧录全部组合兼容;检测不到人脸自动回退居中裁剪——永远不会比以前更差
• 全程本地推理(YuNet,233KB 模型,CPU 每帧毫秒级),素材照旧不上传

📊 画面声音证据(找爆点更准)

• 响度峰值(情绪爆发/笑声)与镜头切换密度本地采集,注入 AI 爆点判断——不再纯文本盲选

📦 发布物料一步到位

• 每条切片自动导出封面 JPG(字幕和标题贴片已烧录,平台直传可用)
• clips.json 元数据:标题/钩子/爆款分/AI 评审意见/时间码/关键词——矩阵运营直接接 CMS

下载

平台	文件
Windows 安装版	HotClip-0.3.0-win-x64.exe
Windows 绿色版	HotClip-0.3.0-win-x64.zip
macOS(Apple 芯片)	HotClip-0.3.0-mac-arm64.dmg

⚠️ 未签名:Windows 点「更多信息 → 仍要运行」;macOS 右键 → 打开。

---

Headline: face-tracking smart reframe. Vertical 9:16 is no longer a dumb center crop — the frame follows the person.

• Face-tracking reframe: shot-aware 3-mode planning (tripod-steady holds / smooth pans / One Euro tracking), fully local (233KB YuNet, ms-per-frame on CPU), graceful center-crop fallback — composes with UI-strip, jump-cut and captions
• Audiovisual evidence: loudness peaks & scene-cut density now feed highlight detection
• Publish-ready assets: per-clip cover JPG (captions & title baked in) + clips.json metadata (title/hook/score/review/timecodes/keywords) for CMS pipelines

v0.2.0 — 从「能用」到「像人剪的」/ From works to hand-edited

这一版的主题:从「能用」到「像人剪的」。 成片质量与全托管体验的大版本。

✨ 新能力

成片质量

• 气口跳剪:自动剪掉说话间的停顿静音并无缝拼接,字幕时间轴同步重映射——成片节奏紧凑,不再有机切感
• 去录屏UI:手机直播录屏的状态栏/固定UI/上下黑边自动检测并裁除(静止的是UI、会动的是内容;检测不到就不裁,普通素材开着无副作用)
• 标题贴片:AI 起的爆款标题自动烧进顶部安全区(黑底白字贴片)
• 字幕三样式:逐字点亮(卡拉OK)/ 智能划重点(AI 选关键词变色放大)/ 大字弹跳(逐块蹦出),一键循环切换;内置思源黑体,任何电脑都不会出豆腐块;字幕基线落在全平台安全区

AI 质量门

• 二阶段复评:严格评审员对每条候选盲评(前3秒钩子/是否断章取义/独立可看性),弱候选透明标记「不建议发布」并默认不勾——不再断章取义
• 关键词、评语、爆款分全部透明展示,可一键否决

转写升级(三档可选)

• 快速 SenseVoice(170MB)/ 均衡 Paraformer(230MB)/ 最准 FireRedASR2(小红书开源,普通话/方言/中英混说,错误率约砍半,520MB)
• 全部本地运行、词级时间戳;无标点模型自动标点回补

全托管

• 一键全自动出片:导入后一个按钮,转写 → 找爆点 → 复评过滤 → 竖屏+字幕+标题+跳剪成片,全程零点击,人只做最后审片

下载

平台	文件
Windows 安装版	HotClip-0.2.0-win-x64.exe
Windows 绿色版	HotClip-0.2.0-win-x64.zip
macOS(Apple 芯片)	HotClip-0.2.0-mac-arm64.dmg

⚠️ 未签名:Windows 点「更多信息 → 仍要运行」;macOS 右键 → 打开。

---

Theme: from "it works" to "looks hand-edited".

• Jump-cut silences: pauses between speech spliced out, captions remapped — tight, human-edited rhythm
• Strip screen-recording UI: status bars / app chrome / letterbox auto-detected (temporal variance) and cropped
• Title card: the AI-written title burned into the top safe zone
• 3 caption styles: karaoke / keyword highlight / word pop — bundled CJK font, platform-safe placement
• Stage-2 AI review: a ruthless reviewer re-judges every candidate (hook, quote-mining, standalone value); weak clips flagged & unchecked by default
• 3 local ASR tiers: SenseVoice / Paraformer / FireRedASR2 (~half the error rate), word-level timestamps, auto punctuation restoration
• One-click hands-off pipeline: import → transcribe → detect → review → export, zero clicks in between

v0.1.0 — 首个公开版本:三步出片 MVP / First public release

HotClip 首个公开版本:长视频 → AI 找爆点 → 竖屏逐字字幕成片,三步全流程可用。

下载

平台	文件
Windows 安装版	HotClip-0.1.0-win-x64.exe
Windows 绿色版(免安装,解压即用)	HotClip-0.1.0-win-x64.zip
macOS(Apple 芯片)	HotClip-0.1.0-mac-arm64.dmg

⚠️ 当前版本未做代码签名:Windows SmartScreen 会提示「更多信息 → 仍要运行」;macOS 首次打开请右键 → 打开(或在「系统设置 → 隐私与安全性」中允许)。

功能

• 导入:MP4 / MKV / MOV / FLV / TS 及纯音频,拖拽即可;直播回放数小时大文件直接进
• 本地转写:SenseVoice 引擎,中/英/日/韩/粤五语种,逐字时间戳;模型(约 170MB)首次自动下载、国内镜像优先,此后完全离线,素材永不上传
• AI 找爆点:接入任意 OpenAI 兼容大模型(推荐 Atlas Cloud,一个 Key 用齐中外主流模型;本地 Ollama 也行)。LLM 只引用原文、不猜时间戳,切点由逐字对齐反推,精确到词;每条候选附爆款分/开场钩子/推荐理由
• 出片:勾选想要的片段,一键切出竖屏 9:16(1080×1920)成片,卡拉OK逐字点亮字幕直接烧进画面;文件落在「影片/HotClip」目录
• 中英双语界面,免费、无水印、不限时长,AGPL-3.0 开源

---

HotClip's first public release: long-form video → AI highlight detection → vertical clips with karaoke captions, end to end.

• Import MP4 / MKV / MOV / FLV / TS or audio-only files — hours-long livestream replays welcome
• Local transcription via SenseVoice (zh/en/ja/ko/yue, word-level timestamps); the ~170MB model auto-downloads once, then everything runs fully offline — your footage never leaves your machine
• AI highlight detection with any OpenAI-compatible LLM (Atlas Cloud recommended; local Ollama works). The LLM quotes text and never guesses timestamps — boundaries come from word-level alignment, so cuts are word-accurate, each with a virality score, hook, and reasoning
• Export selected highlights as 9:16 vertical clips (1080×1920) with burned-in word-by-word karaoke captions
• Bilingual UI (中文/English) · free · no watermark · no length caps · AGPL-3.0

⚠️ Builds are unsigned for now: on Windows, click "More info → Run anyway"; on macOS, right-click → Open on first launch.