World Labs has released Atlas, a world model that builds navigable 3D scenes from multimodal inputs. Meanwhile, Chinese team InSpatio offers an open-source 4D system based on reference video.
Atlas can take text, images, video and 3D information as inputs. From one or several images, it can reconstruct a scene while maintaining spatial continuity as the virtual camera moves.
InSpatio-World follows a different route. It uses video to create a dynamic environment that users can explore from viewpoints not present in the original footage.
Unlike a static reconstruction, the model also represents changes over time. Users can therefore revisit a recorded event from different positions and moments. InSpatio released the model as open source in March.
The Chinese team has also drawn attention through its InSpatio-Curious model. An August 27 interim WorldArena 2.0 leaderboard placed it first among 77 entries with 66.11 points.
Read: AI Bioweapon Risk Prompts RAND Call for Layered Safeguards
It led four measures, including physics adherence, trajectory accuracy and depth accuracy. However, its image-quality score trailed the second-ranked model, showing that the benchmark rewarded more than visual sharpness.
InSpatio is also participating in the Large-Scale Spatial Intelligence 3D Data Open Initiative (SIDO). The programme launched in Wuhan on August 29 with more than 20 universities and institutions.
Read: Google DeepMind’s Genie 3 AI Model Advances Toward AGI with 3D World Creation
SIDO aims to develop large-scale 3D and 4D data, along with common benchmarks, for spatial-intelligence research.
Together, Atlas and InSpatio-World illustrate two approaches to world modelling: reconstructing persistent spaces from multimodal inputs and turning recorded video into explorable environments.
