In brief
Video can be useful for training models that generate video, interpret scenes or act in the physical world. Archives are most licensable when you own the footage outright, people in frame have signed releases or are not identifiable, and you can describe coverage (scenes, actions, duration, resolution) in a structured way.
Why video is in demand
Models that generate video, interpret scenes or control robots need large amounts of footage showing real environments and actions. Footage that is diverse, high-resolution and accompanied by useful metadata tends to be more useful than raw volume alone.
- Stock and production archives with original masters
- First-person or demonstration footage of tasks
- Specialist footage (industrial, medical training, sports, nature) with clear ownership
- Footage with existing annotations, captions or shot lists
Rights and releases
- You (or your company) own the footage, or contributors assigned rights to you
- Model/talent releases cover the people featured, or faces are not identifiable
- Third-party music and logos are absent or cleared
- Prior distribution agreements do not restrict AI training
Note
If you licensed footage from a stock library or another rights holder, read that license before offering it. Standard licenses are written for using footage in your own productions; you should assume you cannot re-license it for AI training unless the terms explicitly say so.
How to describe a video archive
- Total hours and number of clips
- Resolution, frame rate and codecs
- Scene and action categories, with rough proportions
- Capture period and geography
- Annotation status (none, captions, bounding boxes, shot lists)