This paper proposes CamTrol, a training-free and robust method for camera control in off-the-shelf video diffusion models. Unlike previous approaches that require supervised or self-supervised training, CamTrol can be directly integrated with most pretrained video diffusion models to generate camera-controllable videos using a single image or text prompt. The method is inspired by the layout prior that noisy latents in diffusion models influence the final output. By modeling explicit camera movements in 3D point cloud space and using the layout prior of noisy latents, CamTrol enables robust camera control without additional training. The method consists of two stages: first, explicit camera movements are modeled in 3D point cloud representations to generate a series of rendered images indicating specific camera movements. Second, the layout prior of noisy latents is used to guide video generation with camera movements. Extensive experiments demonstrate that CamTrol is effective in controlling camera motion for generated videos and produces impressive results in generating 3D rotation videos with dynamic content. The method is robust and can be applied to various video diffusion models, showing strong generalization ability. CamTrol offers a training-free solution for camera control in video generation, enabling flexible and dynamic video generation without the need for additional training or data annotation.This paper proposes CamTrol, a training-free and robust method for camera control in off-the-shelf video diffusion models. Unlike previous approaches that require supervised or self-supervised training, CamTrol can be directly integrated with most pretrained video diffusion models to generate camera-controllable videos using a single image or text prompt. The method is inspired by the layout prior that noisy latents in diffusion models influence the final output. By modeling explicit camera movements in 3D point cloud space and using the layout prior of noisy latents, CamTrol enables robust camera control without additional training. The method consists of two stages: first, explicit camera movements are modeled in 3D point cloud representations to generate a series of rendered images indicating specific camera movements. Second, the layout prior of noisy latents is used to guide video generation with camera movements. Extensive experiments demonstrate that CamTrol is effective in controlling camera motion for generated videos and produces impressive results in generating 3D rotation videos with dynamic content. The method is robust and can be applied to various video diffusion models, showing strong generalization ability. CamTrol offers a training-free solution for camera control in video generation, enabling flexible and dynamic video generation without the need for additional training or data annotation.