Vorch-Human Unified Multi-Task Human-Centric Generation via Long-Horizon Continuation

Human video generation is advancing rapidly, yet existing systems often solve audio-centered tasks in isolation, limiting cross-task transfer and long-form stability. We present a unified human-centric framework that supports six post-training tasks with a single joint audio-video model. It combines role-aware multi-task conditioning, an automatic pipeline for audio, timbre, identity, and multi-person binding annotations, and a lightweight context-forcing strategy for stable generation up to five minutes. Our model preserves identity and lip synchronization over long sequences while delivering high visual fidelity and strong performance against leading closed-source systems.