李飞飞刚发Atlas，中国开源“同款”已抢跑半年？

TL;DR - World Labs released Atlas, a multimodal world model for spatially consistent 3D scene generation, while Chinese startup InSpatio highlighted its open-source 4D world-model work and benchmark results. The developments signal a shift from plausible video generation toward persistent spatial-temporal environments useful for embodied AI.

• Atlas unifies text, images, video, and 3D context to reconstruct scenes, synthesize novel views, and support controlled camera trajectories.
• InSpatio-World starts from video and explicitly models time, enabling users to observe dynamic events from new viewpoints and moments.
• InSpatio-Curious ranked first in the initial WorldArena 2.0 leaderboard, leading several trajectory, depth, temporal-representation, and physics-adherence metrics, though the benchmark remains limited to specific tasks and distributions.
• The SIDO initiative plans a million-scale open 3D/4D dataset and benchmarks covering reconstruction, generation, scene understanding, spatial reasoning, dynamic prediction, and embodied planning.