Reangle-A-Video: 4D Video Generation as Video-to-Video Translation

We introduce Reangle-A-Video, a unified framework for generating synchronized multi-view videos from a single input video.
We achieve (a) Static view transport and (b) Dynamic camera control without relying on any multi-view generative priors, where both approaches offer six degrees of freedom.

(a) Static view transport: Regeneration of the input video shooting from target viewpoints.

(b) Dynamic camera control: Regeneration of the input video following target camera movements.

Reangle-A-Video decomposes a dynamic 4D scene into view-specific appearance (starting image) and view-invariant motion (image-to-video), addressing each component separately. We first embed the scene's view-invariant motion into a pre-trained video diffusion model using our novel self-supervised training with data augmentation strategy. Initially, to capture diverse perspectives from a single monocular video, we repeatedly perform point-based warping to generate a set of warped videos. These videos, together with the original video, form the training dataset for fine-tuning a pre-trained image-to-video diffusion model with a masked diffusion objective. To achieve (b) dynamic camera control, we sample videos using the fine-tuned model with the original first frame as input. In contrast, for (a) static view transport, we generate view-transported starting images by inpainting the warped first frames under an inference-time view consistency guidance using an off-the-shelf multi-view stereo reconstruction network.

We demonstrate six degrees of freedom in both (a) Static view transport and (b) Dynamic camera control.
Here we visualize the transported viewpoints and camera movements used in our work.

For (a) Static view transport, we compare with Generative Camera Dolly, and Vanilla CogVideoX I2V which employs the same input frame as ours.

For (b) Dynamic camera control, we compare with NVS-Solver and Trajectory Attention.