Valeo fits scaling laws for video diffusion using 5,500 hours of driving footage
abursuc · x · 2026-09-10
VATIX is a valeo.ai research project presented at the ECCV 2026 DriveX workshop: How Far Can 5,500 Hours of Driving Take You? A Scaling Law Analysis of Video Diffusion Models.
Motivation: video generation for autonomous driving can't simply follow web-scale LLM recipes — driving data is expensive, privacy-constrained, and limited in unique coverage, and diffusion models differ from LLMs in training dynamics and scaling behavior. The core question: given a fixed driving dataset, how should compute be allocated to improve generation?
- Data: 5,500 hours of real-world driving from the Natix dataset, covering 28 countries in Europe, North America and Japan, front-facing camera, anonymized faces/plates, 6.3M distinct 2.5s clips at 9 Hz, 320×416.
- Scale: 200+ training runs, 1.6M–1.1B parameters used to fit the laws, with only 3.6% extrapolation error at 9B.
- Three laws: model scaling, training scaling, and compute scaling (allocating model size vs. exposure under a fixed budget), targeting validation diffusion loss via L(x)=L0+A·x^(-α).
Paper, code, models, and dataset are released, with side-by-side ground-truth vs. generated driving clips.
More from Multimodal
- Indie game built with Astra and Higgsfield: full 2D platformer with auto-made sound and music — Critical-Home9648 · 2026-09-10
- Suno V6 vs Google Lyria 3.5: same prompts, side-by-side music generation test — jordiponsdotme · 2026-09-10
- Seedance 2.5 realism demos go viral: "I had to remind myself this is AI" — umesh_ai · 2026-09-10
- Cartwheel MCP + Astra reconstruct moving characters into Blender scenes from video — andrew_n_carr · 2026-09-10
- Minimax H3 tested: 8-step Turbo LoRAs vs 50-step baseline on realistic skin — rm_rf_all_files · 2026-09-10
- Open Actor Protocol Proposed to Standardize Siloed AI Filmmaking Workflows — TheChuckTone · 2026-09-10