PersonaLive is a diffusion-based framework for generating expressive portrait animations from still images suitable for live streaming applications.
The tool addresses the challenge of creating realistic, infinite-length talking-head videos in real time. It uses a streamable diffusion approach that can generate portrait animations while maintaining expressiveness and quality. The framework is designed to work within memory constraints, supporting long video generation on systems with limited VRAM through a streaming strategy for offline inference.
The project is intended for academic research and suits developers working on portrait animation, virtual streaming, or talking-head video generation. Pre-trained weights are available through multiple sources, and the tool integrates with existing ecosystems like ComfyUI. The codebase includes both inference and training implementations, allowing users to either apply the model directly or fine-tune it for specific use cases.
Development activity shows consistent progress with regular releases of core components including inference code, training code, and model weights. The project maintains active engagement with the community through issue resolution and feature enhancements, including WebUI improvements and integration support for third-party tools. A paper describing the approach has been published alongside the code release.
