How the viral-clip picker chooses what to keep
{
"tool": "ViralMint Viral Clip Extractor",
"input": "long-form video (any length, any source) + Whisper transcript",
"selection_axes": ["hook_strength", "retention_curve", "emotion_peaks", "keyword_density"],
"outlier_tiers": {
"OUTLIER": "3x channel median",
"STRONG": "5x channel median",
"BREAKOUT": "10x channel median",
"MONSTER": "20x channel median"
},
"clip_duration_range_seconds": [15, 90],
"output_format": "mp4 + word-by-word ASS captions burned",
"max_clips_per_run": 8,
"aspect_ratios_supported": ["9:16", "16:9", "1:1"]
}
Tier thresholds: outlier_detection_service.py. Clip selector: clip_extraction_service.py. See outlier tiers in the glossary.
Frequently asked
How does ViralMint pick which moments to clip?
After Whisper produces a word-level transcript, an AI scorer evaluates every candidate window against a virality rubric (hook strength, emotional peak, actionable tip, contrarian claim, story loop, number promise, shocking claim, curiosity gap, etc.) and assigns each a 0–10 score. The picker prefers high-score clips while spreading hook types so you don't end up with five curiosity_gaps in a row. Each produced clip carries its score, its hook_type label, and a one-line virality_reason explaining the pick.
What length does the AI pick clips at?
Default is 15–60 seconds — the sweet spot for TikTok / Shorts / Reels. You can override the min and max in the Clip Studio settings (min_duration ≥ 10s, max_duration ≥ 15s, min < max). For podcasts you may want max_duration: 75 to keep a complete argument intact; for fast-paced talking-head content 30 is often enough.
Can it handle landscape source videos?
Yes. Enable force_vertical to convert 16:9 source to 9:16 with blur-fill backdrop on each produced clip. Without that flag, ViralMint preserves the source aspect. Most podcast hosts use 16:9 cameras + want 9:16 output — that's exactly what force_vertical handles.
How long does extraction actually take?
Whisper transcription dominates: ~1 minute of CPU time per 5–8 minutes of source on a typical laptop. After that, clip selection is fast (seconds), and rendering each clip with captions runs ~5–15 seconds per produced clip. End-to-end for a 30-minute podcast producing 10 clips: about 8–12 minutes total. You can walk away — the desktop app pushes a notification when it's done.
Does Clip Studio cost anything per clip?
No. Whisper transcription runs locally, the AI clip picker runs against the cloud chat API (~$0.01 of cloud call per extraction job regardless of how many clips you produce), and rendering uses FFmpeg locally. The only billable parts of ViralMint are the optional paid AI voice (Gemini 3.1 Flash TTS), AI music (Lyria 3 Pro), and AI video clips (Sora 2 Pro / Veo 3.1 / etc.) — none of which Clip Studio uses.
Can I extract clips from a YouTube URL I don't own?
Technically yes — yt-dlp downloads any public video, ViralMint processes it. Whether that's appropriate is up to you: respect the source creator's preferences, fair-use rules in your jurisdiction, and the target platform's rules around third-party clips. For your own content, this is the workflow most creators use to repurpose long-form into shorts.
How does ViralMint compare to OpusClip?
OpusClip is a great SaaS tool with strong AI clip selection. The differences worth knowing: OpusClip is subscription-based ($19–$59/mo with monthly minute caps), watermarks free output, and runs transcription in their cloud. ViralMint's Clip Studio is free in the desktop app with no minute cap, no watermark, no per-clip cost, runs transcription locally on your machine, and the code is AGPL-3.0 open source. Quality-wise both ship strong virality scoring; ViralMint adds hook-type classification (curiosity_gap / contrarian / etc.) and pairs naturally with our Multi-Platform Export for one-click cross-posting.
Is the AI clip picker open source too?
The desktop side (Whisper integration, clip extraction pipeline, caption rendering, FFmpeg orchestration) is AGPL-3.0 at github.com/openclaw-easy/ViralMint. The AI scoring uses an LLM prompt that runs against our cloud chat handler — the prompt itself is in the repo; only the cloud routing is closed.
