Valeo Study: Scaling Laws for Video Diffusion on 5,500 Hours of Driving Data
andrew_n_carr · x · 2026-09-01
Valeo AI releases VATIX, analyzing scaling laws for video diffusion on a fixed driving dataset (5,500 hours, 28 countries).
- Setup: Optimizes model size (N) and training exposure (D) under compute budget (C) to minimize validation loss.
- Key Finding: Training scene exposure reduces loss much faster than parameters (αD ≈ 0.74 vs αN ≈ 0.21), suggesting "seeing more" beats "getting bigger" in data-constrained scenarios.
- Prediction: Laws fitted on 1.1B params predicted 9B model performance with only 3.6% error.
- SOTA: Achieved new state-of-the-art on nuScenes Vista.
Related event: Valeo Derives Video Diffusion Scaling Laws from 5,500 Hours of Driving Data(3 posts)→
More from Research
- Building a long-term memory benchmark for agents: what to add? — True_Mongoose_7073 · 2026-09-01
- Building a high-recall, traceable "second brain" RAG system? — iMiguelmars · 2026-09-01
- Study reveals cross-layer activation patterns in hybrid attention models — 机器之心 · 2026-09-01
- Group-averaged Markov chains papers updated, blending group theory with Markov chains — michaelchchoi · 2026-09-01
- The agent doom loop isn't the model being dumb — it's the transcript working against you — RunAI_Coder · 2026-09-01
- Study: Distilling to Similar Architectures Outperforms Different Ones — SonglinYang4 · 2026-09-01