Abstract
Filmmaking and animation production often require sophisticated techniquesfor coordinating camera transitions and object movements, typically involvinglabor-intensive real-world capturing. Despite advancements in generative AI forvideo creation, achieving precise control over motion for interactive videoasset generation remains challenging. To this end, we propose Image Conductor,a method for precise control of camera transitions and object movements togenerate video assets from a single image. An well-cultivated training strategyis proposed to separate distinct camera and object motion by camera LoRAweights and object LoRA weights. To further address cinematographic variationsfrom ill-posed trajectories, we introduce a camera-free guidance techniqueduring inference, enhancing object movements while eliminating cameratransitions. Additionally, we develop a trajectory-oriented video motion datacuration pipeline for training. Quantitative and qualitative experimentsdemonstrate our method's precision and fine-grained control in generatingmotion-controllable videos from images, advancing the practical application ofinteractive video synthesis. Project webpage available athttps://liyaowei-stu.github.io/project/ImageConductor/