SemCam is a new video-to-video method that lets you control the camera relative to a moving subject using a simple label like "follow" or "maintain frontal view", instead of specifying an exact camera path. The system takes a reference video and a motion label, then creates a new video that shows the requested camera behavior while keeping the subject's identity and action unchanged. It uses a clever shared-basis low-rank adaptation to learn eight different camera motions from limited training data, and a background-consistency loss to keep the scene realistic when new areas come into view.
On a benchmark of 109 scenes, SemCam achieved a 68.6% success rate on semantic motion accuracy, notably higher than the strongest baseline Vista4D at 45.3%, while preserving subject identity just as well. A user study also confirmed the improvement in camera control and visual quality. This makes SemCam the first practical way to control camera behavior at the level of human intent, without needing special trajectory planning tools.