clip_video

io.github.hermoso-ai/hermoso · Verify this server

Cut ONE long video into several RANKED, ready-to-post short clips (podcast, webinar, interview, conference talk, long ad cut → Reels/Shorts/TikTok). Transcribes the source with timestamps, picks the strongest SELF-CONTAINED moments, then cuts + reframes each with ffmpeg — no video model renders anything, which is why it's fast and cheap. THE VERTICAL REFRAME IS SUBJECT-AWARE: a few stills per clip go to ONE cheap vision call, which decides a SINGLE crop offset that is held for that clip's whole length — so a speaker sitting camera-left is not cropped out of their own clip, while the framing still never drifts INSIDE a clip (a per-frame crop truncates to whole pixels and shimmers, so it is deliberately not tracked). It costs one small vision call per clip, billed as its own event. When nothing is being discarded, or no single subject can be located, the crop stays dead centre exactly as before — read `reframedToSubject` and each clip's `reframeWhy` back off the result rather than assuming either way. ACCEPTS: (a) a YouTube link (or Vimeo / Loom / Dailymotion / Streamable / Rumble / Wistia / Twitch / TED) — the server pulls the video down itself; (b) a direct https .mp4/.mov/.webm; (c) a Hermoso /generated/ URL (upload_file turns a local file into one). NOT supported: TikTok / Instagram / Facebook links, and anything age-restricted, private, members-only, geo-blocked or still LIVE — those fail fast with the real reason and are fully refunded, so ask for a direct file or an upload rather than retrying. Source must be at least ~15s and under ~600MB; only the first ~40 minutes is analysed (the result reports truncated:true when it hits that). Cost: a ~7-credit hold, settled to the exact transcription + encode cost, plus the clip-selection model's tokens billed as their own small event. RETURNS clips[] — each with its OWN served mp4 URL, title, hook, ready-to-post caption, 0-100 score and source timecode — not a single video. SUBTITLES ARE BURNED IN BY DEFAULT — slim white CAPS, thin black outline, bottom safe band, no box and no plate — because short-form is watched on mute; pass captions:false for clean footage. TIMING IS APPROXIMATE, NOT WORD-LEVEL: each cue is anchored to the transcript's own per-sentence timestamp and split inside a sentence by character count, so it tracks the speech closely but is not frame-accurate sync — never promise that. Read captionsBurned back off the result: it counts the clips that actually carry a burned track, and captionNote says why any are bare.