Video diffusion models scale effectively with training exposure, study finds

Researchers have conducted a comprehensive scaling-law analysis of video diffusion models specifically for autonomous driving applications. Their study, which involved models ranging from 1 million to 9 billion parameters trained on up to 5,500 hours of driving data, found that increasing training exposure significantly improves model performance more effectively than simply increasing model size under limited compute. However, larger models still achieve lower asymptotic loss, indicating that model size remains crucial for optimal scaling when sufficient resources are available. The study culminated in the training of a 9 billion-parameter model, which reportedly sets a new open-source state-of-the-art for driving video generation on the nuScenes benchmark. AI

IMPACT This research provides crucial insights for optimizing training budgets for video diffusion models in specialized domains like autonomous driving.

RANK_REASON The cluster contains an academic paper detailing a scaling-law analysis of video diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• 1M parameters
• 5,500 hours
• 9B parameters
• Hugging Face
• Natixis Investment Managers
• Nuscenes
• Video Diffusion Models

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →