HunyuanCustom: A New Open-Source Multimodal Custom Video Generation Tool by Tencent Hunyuan 59

What is HunyuanCustom?

HunyuanCustom is a multimodal-driven, customized video generation framework open-sourced by Tencent Hunyuan. It can generate videos featuring specific subjects while supporting multiple input conditions such as text, image, audio, and video.

HunyuanCustom Model Architecture

• Built on HunyuanVideo: HunyuanCustom is built on top of HunyuanVideo and focuses on generating videos consistent with a specific subject.

• Multimodal Fusion: A LLaVA-based text-image fusion module is introduced to enhance the understanding of multimodal inputs. In addition, specific condition injection mechanisms are designed for audio and video to enable video generation under multimodal conditions.

HunyuanCustom Core Features

• Video Generation from Multimodal Inputs: Combines text, image, audio, and video as multimodal inputs to generate videos, delivering a high degree of controllability and generation quality.

• Single-Subject Video Generation: Users only need to upload an image containing the target person or object and provide a text description. The system will identify the identity information in the image and generate coherent, natural video content across entirely different actions, outfits, and scenes. For example, upload a photo of a person, enter "He is walking a dog," and the corresponding video will be generated.

• Multi-Subject Video Generation: Users provide one photo of a person and one of an object, along with a text description, and the two subjects will appear in the video as required. For example, provide a photo of a person and a bag of chips, enter "A man is standing next to a swimming pool, holding a bag of chips and showing them," and the corresponding video will be generated.

• Single-Subject Video Dubbing: In the audio-driven (single-subject) mode, users can upload an image of a person along with an audio clip. The model can then generate effects of the person speaking, singing, or performing other audio-video-synchronized actions in any scene. This is widely applicable to digital human livestreaming, virtual customer service, educational demonstrations, and similar scenarios.

• Video Local Editing: In the video-driven mode, it supports naturally replacing or inserting people or objects from an image into any video clip for creative placement or scene extension, making video reconstruction and content enhancement effortless.

HunyuanCustom Technical Advantages

• High Identity Consistency: Through industry-leading subject consistency modeling capabilities, identity features remain consistent and coherent throughout the video across a variety of scenarios, including single-person, non-human objects, and multi-subject interactions. Characters will not "change faces," and objects will not "drift."

• Strong Controllability: It has strong control signal following capability, including following control signals from multiple modalities such as text, reference subject images, and speech, and can effectively generate the subjects, scenes, and actions users need.

• High Video Generation Quality: Thanks to the powerful capabilities of the Tencent Hunyuan video large model, it achieves industry-leading levels in character detail restoration, motion smoothness, and lighting realism.

HunyuanCustom Application Scenarios

• Advertising Scenarios: Quickly change product backgrounds, and models can rapidly switch outfits.

• E-commerce and Customer Service Scenarios: Quickly and cost-effectively produce vivid digital human product introduction videos, or create digital human customer service videos with specific outfits.

• Film and Television Scenarios: Rapidly produce short dramas and short-story videos.

• Education and Science Popularization Scenarios: Generate explanatory videos through audio-driven generation.

HunyuanCustom's single-subject generation capability has been open-sourced and launched on the Hunyuan official website. Users can experience it under "Model Square - Image-to-Video - Reference-to-Video." Other capabilities have been gradually open-sourced within May. You can access related information through the following links:

• Experience Entry: https://hunyuan.tencent.com/modelSquare/home/play?modelId=192

• Project Website: https://hunyuancustom.github.io/

• Code: https://github.com/Tencent/HunyuanCustom

• Technical Report: https://arxiv.org/pdf/2505.04512