AI clip library
Genyad's clip library is the index that every ad variation is built from. When you upload footage it is split into clips, transcribed, described by what is visibly happening in it, and tagged by type (hook, product, testimonial, B-roll, call to action), so a later request for a variation is a query against your own footage rather than a manual scrub through a folder.
Last updated
What actually gets indexed
Three things per clip, because a script writer needs all three to choose a shot: the words spoken in it, a description of what is on screen, and a role.
- Speech. Every clip is transcribed with word-level timing, which is also what makes burned-in captions line up later.
- Visual context. A short description of the subject and action, so a script can ask for "hands opening the box" and get that clip rather than the one where someone talks about it.
- Clip type. Hook, product, testimonial, B-roll or call to action. This is the field that decides where in an ad a clip is allowed to appear.
- Scope. Which product or campaign a clip belongs to, or nothing at all, which puts it in the shared pool every campaign can draw from.
Indexing runs once per upload. After that the marginal cost of another variation is a credit and about a minute, which is the whole point of the library model.
Why a library beats a timeline for ad variations
A timeline-based editor makes one ad and then makes edited copies of that ad. Two variations built that way share most of their footage, so they fatigue together. A library-based agent selects afresh each time, which is how two variations of the same campaign can end up with no shot in common.
| Timeline editor | Clip library | |
|---|---|---|
| Unit of work | One project file | One indexed clip |
| Second variation | Duplicate and edit | New selection from the pool |
| Overlap between variations | High, usually the same cut | Can be zero |
| Cost of variation 20 | Same as variation 2 | Lower, the library is already built |
| What you search | File names | Speech, on-screen action, clip role |
Creative diversity, not volume, is what slows fatigue. Our own benchmark data on how fast a video ad decays is in the 2026 ad fatigue report.
How to build a library worth querying
- Upload everything you have, including material you rejected. Rejected footage is often the best B-roll.
- Aim for coverage rather than polish: several openings, the product in use, a face talking, and detail shots.
- Leave clips untagged if they are generic. Untagged clips are available to every campaign.
- Tag anything product-specific to that product so a campaign for one SKU cannot pull footage of another.
- Re-upload nothing. Add to the library instead, and every past campaign can use the new material.
Frequently asked questions
Do I need my own footage to use Genyad?
It is the best starting point, but not a hard requirement. If a clip is missing you can generate one from text or an image at 2 credits per second and it joins the library like any other clip. A brand with no video at all is usually better served by an avatar tool first.
Can I search my footage by what is said in it?
Yes. Clips are transcribed on upload, and the script writer selects shots using that transcript plus a description of what is on screen, so a request phrased in plain language resolves to specific clips.
How many clips do I need before the library is useful?
Around 30 usable clips gives enough variety that variations stop reusing the same opening. You can generate ads from far fewer, they will simply overlap more.
Is my footage used to train a model?
No. Your library is scoped to your account and used to assemble your ads.