Video models that can be directed
Runway develops generative video technology and creative tools around it. Its stated research direction includes models that simulate the visual world, while its released products give creators controls over subjects, shots and performances. Those product controls are concrete contributions distinct from the broader research goal. [1]
Gen-4 used reference images with instructions to preserve a character, object or location across different shots. The company demonstrated consistent subjects across lighting and viewpoints, addressing the practical problem of keeping a sequence recognizable instead of treating every clip as a new cast and setting. [2]
Performance and motion are separate controls
Act-One, released inside Gen-3 Alpha in October 2024, used a recorded actor's performance to animate a generated character. A simple camera recording supplied the driving performance, reducing the need for a traditional facial-animation capture and rigging pipeline. [3]
Gen-4.5 followed in December 2025, emphasizing motion and prompt adherence. Its release described additional control modes as forthcoming, so a current feature list should not be projected onto that initial rollout. Selected demonstrations and leaderboard comparisons also do not establish that every generated scene obeys real-world physics. [4]