How Far Can 5,500 Hours of Driving Take You? A Scaling Law Analysis of Video Diffusion Models

TL;DR - A scaling-law study of video diffusion models trained from scratch on up to 5,500 hours of driving data finds that additional training exposure improves loss faster than increasing model size under limited compute. The derived laws guided a 9B-parameter model that achieves a new open-source state of the art for driving-video generation on nuScenes.

• Validation loss follows consistent power laws with both model size and training exposure across models ranging from 1M to 9B parameters.
• For a fixed model and limited compute, longer training is more effective than scaling parameter count.
• Larger models reach lower asymptotic loss, so compute-optimal scaling still favors greater model capacity when sufficient compute and data are available.
• The authors released code and pretrained models, while the underlying driving dataset is being released separately in stages.