Tongyi Open Source Releases Large Model of Ultra-Realistic 3D Digital Human

On April 7, Tongyi Big Model announced the open source release of a hyper-realistic 3D digital human model (LAM). Specifically, it can generate a hyper-realistic 3D digital human model in seconds from a single image: thanks to the 2D image and 3D Gaussian point interaction Transformer designed by Tongyi, it can be trained on a huge amount of video data to obtain a generalizable 3D Gaussian a priori model, which can be used to generate a driveable 3D human model in seconds from a single image. Feedforward generates a 3D human head model that can be driven in seconds; Cross-platform ultra-real-time drive rendering: Tongyi has built the entire drive and rendering framework based on WebGL, and the generated 3D human head model can run in ultra-real-time on any device based on WebGL, and it can run at full 120FPS on the cell phone side; Low-latency real-time interactive and interactive dialog digital human SDK: Based on the generated 3D human, we have built a complete real-time human SDK, which is a real-time interactive and real-time interactive dialog SDK. Based on the generated 3D digital human, we have built a complete real-time interactive dialog engineering link, including LLM, ASR, TTS, and digital human module, i.e., by inputting a picture, we can have a low-latency real-time dialog with the digital human incarnation generated from the picture, which can be applied to customer service, education, companionship, and other scenarios.