PixelAI is an Alibaba-affiliated research and development team focused on visual computing technologies used in digital human creation and enhancement. The team is best known for developing the TaoAvatar system, a high-fidelity, real-time 3D full-body avatar solution based on 3D Gaussian Splatting. PixelAI has released several notable projects through its GitHub page, such as TaoAvatar and GaussianTalker, which support advanced AI-driven avatars with facial expressions, gestures, and real-time speech interaction. Their technologies are designed to work on mobile and AR devices, including the Apple Vision Pro. PixelAI has also developed tools for image enhancement, video restoration, real-time portrait segmentation, and AR-based product interaction (e.g., virtual try-ons), and has won awards in national broadcasting and AI competitions for its innovations in digital human and video processing technologies.
TaoAvatar (taoavatar.org) is a real-time 3D full-body avatar system developed for augmented reality, capable of generating lifelike digital humans from multi-view image sequences. These avatars are fully controllable in terms of pose, gestures, and facial expressions, and run efficiently on mobile and AR devices like the Apple Vision Pro at up to 90 FPS. The technology uses 3D Gaussian Splatting and a personalized parametric template, enhanced through a StyleUnet-based pre-training phase, followed by knowledge distillation into a lightweight MLP network for mobile deployment. The system also includes Audio2BS for natural voice-driven animation. Applications include e-commerce live streaming, holographic communication, and virtual agents. Interactive demos and a high-resolution dataset, TalkBody4D, are available through their HuggingFace repository. TaoAvatar combines photorealism, real-time rendering, and device compatibility, but still faces challenges in fine-grained control and resource demands.