PersonaLive - Real-time AI face-swap live streaming tool open-sourced by the University of Macau and others

What is PersonaLive

PersonaLive is a real-timedigital humanlive streaming tool open-sourced by the University of Macau and GVC Lab. The tool can run on a single consumer-grade GPU with 12GB of VRAM, enabling unlimited-duration portrait animation generation and supporting low-latency interaction. By optimizing VRAM management and TensorRT acceleration, PersonaLive solves the problems of traditional digital human live streaming, such as high hardware barriers, VRAM explosion, and audio-video desynchronization. PersonaLive can easily enable AI streamers for e-commerce and virtual avatars for meetings, bringing new possibilities to the field of digital human live streaming.

Main features of PersonaLive

• Low hardware requirements: Requires only a single consumer-grade GPU with 12GB of VRAM (such as RTX 3060, 4070, etc.), greatly lowering the entry barrier.
• Unlimited-duration generation: Supports continuous video stream generation, suitable for long-duration live streaming, and solves the problems of VRAM explosion and temporal inconsistency in traditional models.
• Low latency and real-time interaction: Through optimized VRAM management and inference acceleration, it supports TensorRT acceleration, achieving audio-video synchronization and instant interaction.
• Ecosystem-friendly: Supports ComfyUI and RTX 50 series GPUs, keeping up with hardware and community trends.
• WebUI visual operation: Provides a simple and easy-to-use web interface, making it convenient for users to upload photos, adjust parameters, and generate animations in real time.

Technical principles of PersonaLive

• Streaming Diffusion framework: Based on the Diffusion Model, it achieves continuous video stream generation through optimized VRAM management and inference acceleration, rather than traditional segmented video generation.
• VRAM optimization: The tool uses efficient VRAM management technology to ensure stable operation over long periods under limited VRAM, avoiding VRAM overflow.
• Temporal module: Introduces a special temporal module that supports unlimited-duration video stream generation, suitable for long-duration live streaming scenarios.
• TensorRT acceleration: Supports converting models to TensorRT format, increasing inference speed by about 2 times and significantly reducing latency.

PersonaLive project address

• GitHub repository：https://github.com/GVCLab/PersonaLive
• HuggingFace model library：https://huggingface.co/huaichang/PersonaLive
• arXiv technical paper：https://arxiv.org/pdf/2512.11253

Application scenarios of PersonaLive

• AI virtual streamers: Used for e-commerce live streaming, news broadcasting, and entertainment live streaming, enabling low-cost, high-quality 24-hour uninterrupted virtual character live streaming.
• Virtual meetings and remote collaboration: Provides virtual avatars in video conferences and remote teaching, protecting privacy and enhancing the interactive experience.
• Content creation and entertainment: The tool can assist content creation such as short videos, films, and games, enriching formats and improving production efficiency.
• Customer service and support: Supports building virtual customer service and intelligent assistants, enabling 24-hour online service and improving customer experience.
• Education and training: Generates virtual teachers and training simulation scenarios, providing personalized and efficient teaching and training services.