世界模型ICML 20252025
Sora Turbo: High-Fidelity Video Generation with Physics Understanding
European Strategy for Particle Physics Preparatory GroupOpenAI
摘要
We present Sora Turbo, an enhanced video generation model with deep physics understanding. Sora Turbo generates photorealistic videos that respect physical laws including gravity, collisions, and fluid dynamics, enabling accurate simulation of complex real-world scenarios from text descriptions.
video generationphysics simulationtext-to-videophotorealisticphysical laws
技术细节
数据集
Internal Video Dataset (web-scale)Physics-101
仿真平台
Large-Scale GPU Cluster
模型骨架
Diffusion Transformer (DiT) + Spacetime Patches
编码器
ViT Spacetime Patch Encoder
解码器
DiT Decoder (28 layers) + Video VAE Decoder