Fei-Fei Li Just Released a Landmark Model - A Chinese Open-Source Team Did It Six Months Ago

Fei-Fei Li recently unveiled a groundbreaking model, but a Chinese team achieved a similar accomplishment six months prior. Atlas, Li's creation, is the first omni-world model that integrates text, images, video, and 3D data into a unified space. It can generate a 3D scene from a single image and allows users to explore it as if they were physically moving through the space.

Meanwhile, Yingsu, a lesser-known Chinese company, released an open-source model called InSpatio-World in March. This model can turn video footage into a dynamic 4D scene, enabling users to alter both the perspective and the time of the event. The unique aspect of InSpatio-World is its ability to reconstruct the space even if the filmed moment was never captured.

The two models serve different purposes: Atlas builds a world from static images, while InSpatio-World records the real world by processing video footage. Notably, Yingsu's upgraded version of InSpatio-World will also be open-source, mirroring ImageNet's impact on the deep learning landscape.

A WorldArena 2.0 benchmark tested the practical utility of embodied world models by evaluating robotic arm trajectories, depth stability, and object persistence. Yingsu's InSpatio-Curious model emerged as the victor, excelling in physics adherence, trajectory accuracy, JEPA similarity, and depth accuracy. However, its image quality score was comparatively lower than its closest competitor.

The significance of these developments lies in the emergence of a robust spatial intelligence data program called SIDO. Launched by Yingsu and over 20 universities, SIDO aims to create a million-scale 3D/4D scene database, fostering a comprehensive understanding of the world for world models. This partnership between American and Chinese teams highlights a shared goal in advancing artificial intelligence, with Atlas representing the Western approach and InSpatio-World embodying the Chinese perspective.