Researchers have released CamCloneMaster, a framework that lets you control camera movement in AI-generated video by showing it a reference clip rather than specifying camera parameters.
According to the project page, you feed it a video with the camera movement you want, plus optional content reference footage, and it replicates that motion in the generated output.
What makes it different:
Works without requiring explicit camera parameters or test-time fine-tuning
Handles both image-to-video and video-to-video generation in one framework
Uses a 3D VAE encoder to convert reference videos into conditional latents that guide the generation
The training approach: The team built a synthetic Camera Clone Dataset using Unreal Engine 5, combining 3D scenes with character animations and multiple paired camera trajectories. They acknowledge that building this kind of triple-video dataset (camera reference, content reference, and target) in the real world would be difficult and labor-intensive.
The research includes user studies showing improvements in both camera controllability and visual quality compared to existing methods, though specific availability details aren't provided.
Worth noting: Camera control remains one of the trickier aspects of AI video generation. If the "show rather than tell" approach holds up in practice, it could be genuinely more intuitive than current parameter-based methods—assuming this research eventually makes it into production tools.
