Morphic publishes Reshoot-Anything, a method that changes the camera angle of an existing video.

Morphic releases Reshoot-Anything, an AI method by Adithya Iyer that changes camera angles in existing videos using dynamic crops to train on new perspectives.

Adithya Iyer and collaborators release the code and paper for Reshoot-Anything, a method for generating new perspectives from an existing video. The topic is central to AI video generation: for a model to learn to modify a scene's angle, it ideally needs to be trained on multi-view videos, i.e., simultaneous shots of the same event from multiple cameras. However, these videos are rare in nature, and available synthetic datasets lack diversity.

The paper's main contribution is to circumvent this scarcity. Rather than searching for real multi-view videos, the authors sample different dynamic crops within the same video: a random crop serves as the source, another is predicted as output, which amounts to training the model to re-render a scene from a different viewpoint. Combined with a small amount of synthetic videos, this approach is sufficient to make the model generalize to complex camera trajectories (orbits, pans) without degrading the visual quality of the base model. Simple augmentations (Gaussian noise, image warping, fluorescent backgrounds on anchor frames) strengthen the model's robustness against 4D projection errors.

The rendering, illustrated in several demos, handles these complex camera movements while preserving the dynamics of the source video. The model weights, code, and paper are publicly available on arXiv and on Morphic's GitHub.