Zovda
A fifteen-second clip carries its whole point in the first two seconds. A forty-minute recording carries it somewhere in the middle. Same file format, completely different training problem.
Teams that reuse a long-video pipeline for short clips usually end up with labels that describe an average of the clip, which is exactly what no viewer ever sees.
The hook is a label
For short video, the opening is the most valuable segment and the least annotated. What appears first, how fast the cut lands, whether text is on screen: those facts decide whether the content works.
A corpus that labels only the middle of the clip throws away the part that carries the behaviour.
Segment at cuts, not at seconds
Fixed-interval sampling slices through cuts and produces frames that never existed as a shot. Cut-aware segmentation keeps each shot whole and records its length.
That single choice changes what a model can learn about pacing.
Metadata carries half the signal
Duration, aspect ratio, shot count, on-screen text presence, audio type, and a content category. Six fields let a buyer build a set for "fast cuts with captions" without decoding a single file.
We keep them as columns, not as prose in a manifest.
The checklist we run
- Are segments cut at shot boundaries rather than fixed intervals?
- Is the opening seconds range labelled separately?
- Is on-screen text recorded as a field?
- Can clips be filtered by pacing without decoding?
- Do short and long families stay separate in the catalogue?
If a supplier cannot answer these, short video is just long video with fewer frames. At Zovda, the Datuva catalogue keeps the two families apart because the training signals differ.
Where to start
Start from the behaviour you want to change: watch time, completion, or first-frame attention. The video collection separates pacing-driven and procedure-driven material, and the corpus catalogue shows the stage mapping.
Tell us the behaviour and the volume. We reply within one business day.