HunyuanCustom is a multimodal conditional video-generation pipeline designed to preserve subject identity while supporting flexible conditioning from text, images, audio, and video. Built on a HunyuanVideo backbone, it introduces modules to close modality gaps and prevent identity drift so generated videos remain faithful to a provided subject image(s). Key components include a Text–image fusion module (LLaVA-based) for injecting image identity cues into text prompts, an Image ID enhancement module that temporally concatenates image features across frames, AudioNet for audio-conditioned alignment of audio and visual features, and a patchify-based video-driven injection for latent-conditioned editing. The framework uses a disentangled identity representation to decouple identity information from other modalities, enabling independent control over image, audio, and video inputs. Supported workflows include single-subject video customization, video-driven editing with masks, and audio-driven customization, with downstream use cases such as singing avatars and virtual advertisements. Deployment details cover Linux, CUDA, PyTorch, multi-GPU setups, and Docker/Gradio-based tooling.