TeleMorpher Does the Impossible: Edits Video Motion and Location in One Shot

Recent diffusion models excel at generating and editing images and videos, but editing both the motion and location of a subject simultaneously remains largely unexplored. Haengbok Chung's TeleMorpher addresses this gap with one of the first one-shot frameworks for robust simultaneous motion-location editing. The method first disentangles the protagonist from the background using pre-trained segmentation and inpainting models. It then employs a training-free pose warping step guided by motion priors generated from an off-the-shelf model, injecting the warped motion directly into a baseline motion editor during inference to preserve source appearance while matching target motion.

To evaluate performance, TeleMorpher introduces two new LPIPS-based metrics: one measures background consistency before and after editing, and the other quantifies motion fidelity by comparing protagonist skeletons between source and target videos. Experiments on the TaiChi dataset and real-world videos demonstrate that TeleMorpher outperforms existing methods in both quantitative and qualitative human evaluations. This work promises to unlock practical applications in video production, animation, and AR/VR by allowing seamless edits to both where a subject is and how it moves, all without additional training.