Complete collection of generated videos from our method
Featured examples demonstrating subject-driven video generation with diverse prompts.
Complete collection of videos generated by our method across various subjects.
Qualitative results on VBench and Open-S2V benchmarks comparing with tuning-free baselines.
Comparison with tuning-based methods (CustomCrafter, Still-Moving) using identical reference images and prompts.
While our training data contains limited human faces, the method still shows reasonable generalization.
Visualizing the effect of different training configurations.
Our training dataset (Subject-200K) primarily comprises general objects with few human faces, which may hinder effectiveness for human-specific personalization.