In the field of digital humans, Tsinghua Shenzhen Institute and the Chinese team of the International Institute for Digital Economy recently launched a new technology called GUAVA, marking a new era in the creation of digital humans. With just one photo, GUAVA can generate a high-quality 3D Gaussian avatar in 0.1 seconds and drive it in real-time, with a frame rate exceeding 50 frames per second.
Traditionally, creating a high-quality 3D digital human required complex multi-view shooting or time-consuming video data training, often taking several hours to complete. The emergence of GUAVA undoubtedly disrupts this process. Compared to other methods, GUAVA's reconstruction speed is astonishing, taking only 0.1 seconds, while other algorithms such as ExAvatar take 2.4 hours, GaussianAvatar takes 1.3 hours, and even GART requires 7 minutes. This speed has amazed people.
GUAVA's superior performance is due to its innovative technical architecture, which mainly includes two key components: EHM model and 3D Gaussian splatting. The EHM model ensures high-fidelity and precise control of facial expressions by combining SMPLX and FLAME technologies, while 3D Gaussian splatting achieves fast rendering by splitting the scene into millions of 3D Gaussian spheres. This method makes GUAVA also perform well in identity consistency, with relevant indicators exceeding those of competitors.
In practical applications, GUAVA can provide strong support for multiple fields such as self-media, live streaming, e-commerce, and education. Self-media users can quickly create customizable characters with just one image, greatly shortening the production cycle; live streamers can easily upload selfies and quickly convert them into virtual avatars; e-commerce platforms can achieve virtual models tailored for each individual; the education industry can also use virtual teachers for immersive teaching. These application scenarios demonstrate the wide potential of GUAVA technology.
Tsinghua Shenzhen Institute and IDEA Institute have demonstrated their leadership in the field of digital human technology through actual achievements without relying on funding or hype. Their research has been recognized at the ICCV 2025 conference and has made this innovative achievement accessible to the global community through open-source code.
References:
https://github.com/Pixel-Talk/GUAVA
https://eastbeanzhang.github.io/GUAVA/
Key Points:
- 🚀 GUAVA technology can generate a 3D digital human from a single photo in 0.1 seconds, which is astonishingly fast.
- 🎨 Its core technologies, EHM model and 3D Gaussian splatting, ensure high-quality expression restoration and fast rendering.
- 💡 GUAVA is widely applied in multiple fields such as self-media, live streaming, e-commerce, and education, improving efficiency and user experience.
