HunyuanCustom

HunyuanCustom is designed to address the core challenges of customized video generation: maintaining subject identity and supporting diverse input modalities. Building on the HunyuanVideo framework, it introduces a novel image-text fusion module (LLaVA) for richer multimodal understanding, and an identity enhancement mechanism that leverages temporal modeling to keep subjects consistent across frames. For audio and video-driven scenarios, HunyuanCustom employs specialized condition injection networks, enabling precise and disentangled control over each modality.

Extensive experiments show that HunyuanCustom not only excels in single- and multi-subject video generation, but also achieves state-of-the-art performance in realism, identity preservation, and flexible scenario adaptation.