The latest RTFM (Read The Field Model) introduced by Li Fei-fei's team has become one of the most groundbreaking 3D world generation models. This model can achieve real-time inference at interactive frame rates on a single NVIDIA H100 GPU, successfully advancing "3D world generation" from concept to practical application.

RTFM's biggest highlight is its ability to run in real-time with persistence and 3D consistency. The model not only generates complete 3D scenes but also maintains stable geometric structures, object positions, and appearances during interactions, supporting complex visual effects such as reflections, shadows, specular highlights, and glows, with realism comparable to game engines.

Different from previous short-term generated 3D models, RTFM introduces a "persistent memory mechanism," enabling the generated virtual world to have long-term continuity. Users can interact and explore the 3D space created by the model for an unlimited duration, and the scene will not disappear due to changes in viewpoint or actions, truly realizing an AI world that can exist sustainably.

Industry experts believe that the release of RTFM marks a critical step forward for AI world models (World Model) toward high-fidelity real-time rendering, providing a new infrastructure for fields such as virtual reality, game engines, and embodied intelligence in robotics.