One-Shot Visual Identity Transfer in Dynamic Scenes Using Target Images

Abstract

Visual identity transfer in dynamic scenes refers to the process of generating a video in which the motion, viewpoint, and environment of a source sequence are preserved while the appearance of the central subject is replaced by the identity of a target individual. Applications include personalized content creation, privacy-aware editing, and virtual production, where a single image of a new identity is often the only available reference. This one-shot regime is challenging because the model must infer viewpoint- and illumination-consistent identity features from incomplete observations while maintaining temporal coherence under arbitrary motion. Existing approaches typically rely on multi-view capture, subject-specific fine-tuning, or strong priors that can limit fidelity or generalization. This work studies a general formulation of one-shot visual identity transfer using target images in unconstrained dynamic scenes and introduces a model that separates identity from scene dynamics within a conditional generative framework. A single target image is mapped to a compact identity embedding that drives a spatiotemporal generator conditioned on a source video. The method combines appearance modeling, motion-aware warping, and temporally consistent adversarial training to synthesize videos that preserve source motion and background while approximating the target identity. The study examines the impact of architectural choices, training objectives, and numerical optimization schemes on identity preservation and scene consistency, and analyzes failure modes such as extreme pose mismatch and occlusions. The resulting formulation highlights the interaction between representation, training dynamics, and generalization in one-shot identity transfer for dynamic visual content.