GAIA-2:
A Controllable Multi-View Generative World Model for Autonomous Driving

GAIA-2 is a latent world model trained with flow matching in the continuous latent space induced by a video tokenizer. It allows scalable simulation of diverse driving scenarios, reducing the reliance on expensive real-world data collection and facilitating robust evaluation in safe and repeatable environments. In particular, our world model excels at generation of edge-case/safety-critical data, resimulation of existing scenarios, and extreme generalisation.

Our model can generate novel driving scenarios with remarkable multi-view and temporal consistency. Compared to its predecessor, controllability has been significantly improved through fine-grained control on ego-vehicle actions, dynamic agents behaviour, scene geometry, and environmental factors.