Accepted at the 6th DriveX Workshop in conjunction with ECCV 2026
This is the official repository for the paper "How Far Can 5,500 Hours of Driving Take You? A Scaling Law Analysis of Video Diffusion Models".
Authors: Victor Besnier, Anh-Quan Cao, Elias Ramzi, Spyros Gidaris, Tuan-Hung Vu, Andrei Bursuc, Eloi Zablocki, and Matthieu Cord. Affiliation: Valeo, Valeo.ai, Paris.
The code and pretrained models for the VATIX model family will be publicly released shortly. Please star or watch this repository for updates.
Video generation for autonomous driving cannot follow the web-scale route: driving data is expensive to collect, bound by privacy requirements, and cannot be scraped at will, meaning models must make the most of a fixed corpus.
We present a systematic scaling-law study of video diffusion models trained from scratch on driving data. We trained a family of Diffusion Transformer (DiT) flow-matching models ranging from 1.6M to 9B parameters, using up to 5,500 hours of driving data.
- Consistent Scaling Laws: Validation loss follows consistent power laws in both model size and training exposure.
- Compute Optimization: Loss improves much faster with training exposure than with model size, making longer training the most effective way to improve a fixed model under limited compute.
- The Value of Scale: Larger models continue to achieve lower asymptotic loss, meaning compute-optimal scaling still favors increasing model size when sufficient compute and data are available.