HunyuanCustom

Released May 8, 2025, HunyuanCustom is Tencent's open-source framework for customized video generation, built on the HunyuanVideo-13B backbone. Where most video models take a single prompt, HunyuanCustom accepts text, single or multiple images, audio, and even video as joint conditioning: give it reference images and it preserves subject identity across generated clips; give it audio and the subject speaks or sings it; give it a video and it can swap in a subject from an image. The team reported top scores on face similarity and subject consistency against contemporaries like VACE, SkyReels, Pika, Vidu, and Hailuo.

Tencent open-sourced the inference code and weights progressively — single-subject first, then audio-driven and video-driven variants — and the community quickly integrated it into ComfyUI, with single-GPU 8GB VRAM support arriving via community tooling within weeks.

Limitations: as a 13B-class video model it needs meaningful GPU memory for quality output, and multi-subject and fine-grained motion control remain weaker than single-subject identity preservation. Customization models are dual-use by nature — identity-preserving generation demands consent-aware deployment.

Vendor-reported: #1 face similarity and DINO similarity vs VACE, SkyReels, Pika, Vidu, Hailuo