The video Motion Prompting research by Google DeepMind controls video generation through motion trajectories, with the core objective of designing a generative system that uses motion information to generate dynamic video content in a more flexible and controllable manner.
Train trajectory-based conditional video generation models.
to represent motion: This flexible motion representation supports encoding single or multiple point trajectories. It can describe the motion of specific objects or the entire scene. Including occluded and temporally sparse motion sequences.
Method steps
Train trajectory-based conditional video generation models. 
Use motion prompts to guide the model to generate target behaviors. 