HunyuanCustom

HunyuanCustom generates videos of specific people or objects for creators and researchers who need subjects to stay recognizable across frames. It builds on HunyuanVideo and supports generation on your own hardware.

• Reference-based generation: Use an image to define the subject and a text prompt to describe the scene and action.
• Speaking and singing avatars: Pair a reference image with supplied audio to generate a character speaking or singing that recording.
• Subject replacement: Replace a selected subject in an existing video with a person or object from a reference image.
• Browser and ComfyUI access: A Gradio interface provides access to a self-hosted instance, while ComfyUI-HunyuanVideoWrapper supports image-based subject customization; its integration doesn't cover the audio and video workflows.

Local inference requires an NVIDIA GPU with CUDA support. Linux is the tested platform, and the code supports single or multiple GPUs. The standard path recommends 80GB of GPU memory; its lower-memory 720p path needs at least 24GB and runs very slowly. A separate low-memory route supports a single GPU with 8GB of VRAM.

Downloadable model weights and inference code let you run generation on your own machine or server. Cloud-Native-Build (CNB) also offers a cloud route, where generation runs on hosted hardware.