GAIA-1

GAIA-1 is a generative world model developed by the UK company Wayve for autonomous driving. It takes video, text and action (vehicle control) inputs and generates realistic driver-perspective video sequences that are physically and geometrically consistent with the driving scenario.

Architecture

The model combines an autoregressive transformer (~6.5B parameters) operating on discrete video, text and action tokens with a diffusion video decoder (~2.6B parameters) that renders continuous frames from those tokens. Roughly 9B parameters in total. Trained on ~4,700 hours of proprietary driving data collected by Wayve in the United Kingdom.

Use cases

GAIA-1 does not control the vehicle — it is used to generate synthetic data and scenarios for training and evaluating autonomous driving stacks, including rare corner cases. Weather, lighting, behaviour of other road users and ego-vehicle commands can be controlled via text prompts and action vectors.