100,000 hours in under a week for one of the largest video data providers
VideoDB checked more than 100,000 hours of footage and cut it into scene-level training samples for one of the largest video data providers, in under a week, with the files left in the provider's storage.
One of the largest video data providers licenses footage to teams that train video generation models. VideoDB checked more than 100,000 hours of its archive and cut it into scene-level samples in under a week. The files never left the provider’s own storage.
| Customer | One of the largest video data providers |
| Footage | More than 100,000 hours, from many sources |
| Delivered | Scene-level samples for training video generation models |
| Time | Under a week |
| Files | Left in place, in the provider’s storage |
The problem
Teams that train video generation models buy samples: clean clips the length of a scene, with enough variety to teach a model something new. The provider’s archive came from many sources, in many formats, and nobody had a reliable record of what each file held. Some files were damaged, some shots were static or blurred, and the same footage turned up more than once.
Turning that archive into samples meant reading every frame of it, and at 100,000 hours the work had to run where the files already were.
What VideoDB did
- Read the archive where it was. VideoDB connected to the provider’s storage and worked on the files in place, writing its results as a layer on top of them. Nothing was copied out.
- Checked every clip. Each file was checked for video health (corrupted files, broken keyframes, frame rate and resolution), visual validity (black, blurred, dark or static frames) and camera stability, and near-duplicates were found across sources.
- Cut scene-level samples. Long files were split into scenes, each with timestamps that point back to its source file, so every sample can be traced and replaced.
- Spread the work out. The pipeline ran across as many machines as the job needed, which is how the whole archive finished in under a week.
The result
| Hours processed | More than 100,000 |
| Time | Under a week |
| Footage moved | None. The files stayed in the provider’s storage |
| Ready for | Millions of hours |
The provider could tell a buyer what its archive held, scene by scene, and hand over samples that had already passed the checks.
The same pipeline, for any data provider
Today the pipeline also measures variety (which scenes, objects, embodiments and views a dataset covers) and scores actions, with action jumps and state alignment for robot data. A checked dataset can then become an MCP server that researchers’ agents search in Claude or Cursor, the way more than 50,000 hours of public robot and egocentric video are searchable on datasets.bot.
If you sell video data, send us a few clips from each source and we will send back a quality sheet for each one. Talk to us, or see how quality checks for data providers work.