Runway's Gen-3 Alpha announcement connects control over a video's progression to how the model was trained. The company says it used images and videos alongside detailed captions describing scenes over time. Those captions are intended to help the model follow directions about transitions and the timing of changes within a shot. The June 2024 announcement presents this as a route toward more precise temporal control.
For a creator, the relevant capability is directing how a scene unfolds, including camera movement and transitions, through written descriptions. That is the stated training objective behind the release's emphasis on control.
Terms in this piece · Glossary
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
