Starts mid-sentence
A 90 with the first word clipped is a 40. Always check the in-point; nudge back a second.
An AI viral moment finder scans a long video and ranks the segments most likely to work as a short. It reads the transcript for hooks and payoffs, listens for laughter and emphasis, watches for scene changes and reactions, then scores each candidate. This page explains what the scores measure, where they fail, how the main tools differ, and gives you a hand scorer that uses the same five criteria so you can check a tool's picks, or make your own.
It slides a window across the video, builds features for each window from the transcript, audio, frames and (for streams) chat, and scores them with a model trained on clips that did and didn't hold attention. The output is a ranked list with timestamps, usually labelled a Virality Score (OpusClip, Klap), a multi-axis score (Katto's four axes) or just "highlights" (Vizard, Submagic).
Different tools weight them differently; nearly all of them use these five.
Does the first line create a question, a claim or a contradiction? "I lost $40,000 in a week" scores; "so, um, yeah" doesn't. Tools look at the opening sentence's structure and at whether the window starts mid-thought. This is the biggest single weight in most models.
Laughter, raised volume, speed changes, a pause before the answer, strong sentiment words. Audio features catch what the transcript misses: a joke's delivery, a gasp, disbelief.
Can a stranger follow it with zero context? Windows full of "like I said earlier" or unexplained names score low. Good tools nudge the boundaries to start where the idea starts.
Is there a resolution: the punchline, the number, the reveal, the reaction? Clips that end on the answer get rewatched and shared; clips that fade out mid-point get swiped. Models detect sentence-final emphasis and tonal closure.
Money, relationships, controversy, "nobody tells you" framing, numbers, named entities people search for. A tight, repeatable line ("rent is a scam") doubles as caption and hook. Prompt-first tools (Katto) let you steer this axis directly.
Rate a candidate on the five signals. The weights mirror what the commercial scorers do: hook and payoff count most. Use it to sanity-check a tool's pick or to rank your own cuts.
—
Scoring: weighted sum × 10, minus 1.5 points per second beyond 60 s and minus 2 per second under 15 s, floored at 0. Not any vendor's formula; a reasonable stand-in for all of them.
| Tool | Score shown | Inputs | Steerable? | Best on | From | Review |
|---|---|---|---|---|---|---|
| OpusClip | Virality Score 0–100 + reason | Transcript, audio, visual (ClipAnything) | Prompt per video, keywords | Anything, incl. no-dialogue footage | $15 · Pro $29 | OpusClip |
| Klap | Virality Score per clip | Transcript-led, some visual | Topic hints, length | YouTube and gaming, with Reframe 2 | $29 · $14 annual | Klap |
| Katto | 4-axis score (hook, engagement, clarity, shareability) | Transcript + prompt | Prompt-first, re-run free | Niche content you can describe | €8.90 · €4.90 annual | Katto |
| Vizard | Highlights, ranked | Transcript | Length, topic | Podcasts and interviews | $29 · $14.50 annual | Vizard |
| Submagic | Magic Clips, ranked | Transcript | Minimal | Talking-head, caption-first | $19 · add-on | Submagic |
| StreamClipping | Highlights, up to 30/stream | Chat, audio, visual, transcript | Preferences | Twitch / Kick / YouTube streams | €19 | StreamClipping |
| SupoClip | LLM-ranked segments | Transcript, your model | Edit the prompt | Private, bulk | $0 self-hosted | SupoClip |
A standalone "moment finder" without clipping is rare; every tool above bundles finding with cutting. Find best video moments covers the manual route.
A 90 with the first word clipped is a 40. Always check the in-point; nudge back a second.
Models tolerate 60–90 s windows; platforms don't. A high score on a long clip means "contains a great 30 seconds".
Audio features over-reward shouting. Gaming and reaction content needs the visual and chat signals to balance it.
Hot takes score high and can be the clip that gets you (or the streamer you clip for) in trouble. Campaign briefs ban some; read them.
The model was trained on everyone's clips. Your niche's "viral" may be a 70 that the model never surfaces; use a prompt or the hand scorer.
Top candidates often overlap the same two minutes. Dedupe before you post five versions of one joke.
Set the tool to 15–20 per hour. The ranking is noisy around the cut-off; the gems are often ranked 6–12.
Read the first line of each. If it doesn't earn the next three seconds, skip no matter the score.
In-point at the start of the idea, out-point on the payoff. This moves real performance more than captions or music.
Spacing clips from one source lets you see which hooks worked before the next batch.
Note which scores matched retention. Adjust the tool's prompt or your own weights in the scorer above; after a few batches the finder is tuned to your audience, not the internet's.
It's a useful sort, not a forecast. In practice the top quartile of scored clips outperform the bottom quartile reliably, but within the top ten the order is close to noise. Review, don't trust.
OpusClip's ClipAnything for mixed or visual footage, Katto when you can describe what you want, StreamClipping for live streams, Vizard for pure talk. For most creators the differences between picks are smaller than the difference a good in-point makes.
Yes: SupoClip (open source) transcribes and asks an LLM to rank segments; you can edit its prompt to change the criteria. Expect weaker results on non-dialogue footage.
No. The score is about the moment. Captions and hook titles help retention after the pick; they aren't part of the ranking.
Indirectly. Look at which of the five signals your top clips win on and which they lack; most creators find "payoff" is the gap, which is a scripting fix, not an editing one.