The field of video generation is moving towards incorporating physical laws and principles to improve the realism and coherence of generated videos. Recent works have focused on developing frameworks that can enforce Newtonian mechanics, such as constant-acceleration dynamics and mass conservation, to generate more physically plausible videos. Another direction is to improve the evaluation of video generation models, with benchmarks that assess their ability to reason about physical phenomena and generate videos that are consistent with scientific laws. Noteworthy papers include Post-Training Newton's Laws with Verifiable Rewards, which proposes a physics-grounded post-training framework for video generation, and PhyVLLM, which incorporates physical motion modeling into video language models. These advancements have the potential to significantly improve the quality and realism of generated videos, and to enable more accurate modeling of real-world phenomena.
Physics-Informed Video Generation
Sources
What about gravity in video generation? Post-Training Newton's Laws with Verifiable Rewards
RealGen: Photorealistic Text-to-Image Generation via Detector-Guided Rewards
Seeing the Wind from a Falling Leaf
Envision: Benchmarking Unified Understanding & Generation for Causal World Process Insights
PhyDetEx: Detecting and Explaining the Physical Plausibility of T2V Models
PAI-Bench: A Comprehensive Benchmark For Physical AI
Objects in Generated Videos Are Slower Than They Appear: Models Suffer Sub-Earth Gravity and Don't Know Galileo's Principle...for now
FineGRAIN: Evaluating Failure Modes of Text-to-Image Models with Vision Language Model Judges
RULER-Bench: Probing Rule-based Reasoning Abilities of Next-level Video Generation Models for Vision Foundation Intelligence
Benchmarking Scientific Understanding and Reasoning for Video Generation using VideoScience-Bench
Rethinking Prompt Design for Inference-time Scaling in Text-to-Visual Generation
MoReGen: Multi-Agent Motion-Reasoning Engine for Code-based Text-to-Video Synthesis
PhyVLLM: Physics-Guided Video Language Model with Motion-Appearance Disentanglement
DraCo: Draft as CoT for Text-to-Image Preview and Rare Concept Generation