NEWTON: Agentic Planning for Physically Grounded Video Generation

Researchers introduced NEWTON, an agentic planning framework for physically grounded video generation.

The paper argues that text prompts alone often fail to specify the physical dynamics needed for realistic videos.

NEWTON combines physics-aware tools such as keyframe generation and scientific computation, then verifies and replans outputs.

It reports higher VideoPhy-2 accuracy on LTX-Video and Veo-3.1 without changing the base generator.