WildActor: Unconstrained Identity-Preserving Video Generation

Production-ready human video generation requires digital actors to maintain strictly consistent full-body identities across dynamic shots, viewpoints and motions, a setting that remains challenging for existing methods. Prior methods often suffer from face-centric behavior that neglects body-level consistency, or produce copy-paste artifacts where subjects appear rigid due to pose locking. We present Actor-18M, a large-scale human video dataset designed to capture identity consistency under unconstrained viewpoints and environments. Actor-18M comprises 1.6M videos with 18M corresponding human images, covering both arbitrary views and canonical three-view representations. Leveraging Actor-18M, we propose WildActor, a framework for any-view conditioned human video generation. We introduce an Asymmetric Identity-Preserving Attention mechanism coupled with a Viewpoint-Adaptive Monte Carlo Sampling strategy that iteratively re-weights reference conditions by marginal utility for balanced manifold coverage. Evaluated on the proposed Actor-Bench, WildActor consistently preserves body identity under diverse shot compositions, large viewpoint transitions, and substantial motions, surpassing existing methods in these challenging settings.

This paper studies how to generate videos of a person while keeping their identity consistent across different scenes, camera views, distances, and motions. Existing video generation systems can produce realistic clips, but they often change a person’s face, body shape, clothing details, or overall appearance when the viewpoint or action changes. To address this, we build a large dataset of human videos with multiple views of the same person, and use it to train a system called WildActor. WildActor learns to use several reference images of a person so that the generated videos preserve both facial and full-body identity while still allowing flexible motion and diverse environments. Experiments show that WildActor produces more consistent human videos than existing methods, especially in challenging cases with large viewpoint changes and long multi-shot narratives.