GAIA-2

GAIA-2 is a controllable, multi-view generative world model developed by Wayve for autonomous driving, introduced on 26 March 2025. It is described in the paper “GAIA-2: A Controllable Multi-View Generative World Model for Autonomous Driving” (arXiv:2503.20523). The model generates realistic, temporally and spatially consistent multi-camera video for simulation, scenario testing and augmenting training data.

Its architecture consists of a video tokenizer that compresses sequences into a continuous latent space and a latent-diffusion world model that predicts future scene states. Unlike GAIA-1 (autoregressive), GAIA-2 uses latent diffusion, natively supports multiple cameras and eliminates temporal discontinuities, improving motion smoothness and cross-view coherence.

The model accepts rich structured conditioning: ego-vehicle actions (speed, steering curvature), dynamic agent behaviors (from 3D bounding boxes), road attributes (number of lanes, speed limits, crossings, traffic lights, intersections), environmental factors (weather, time of day) and CLIP plus proprietary model embeddings. It covers data from the UK, US and Germany.

GAIA-2 is used to generate synthetic scenarios (including rare and safety-critical ones), test out-of-distribution generalization and augment real-world data. It is a research, proprietary model (weights not publicly available).