[CVPR 2026] Chain of Event-Centric Causal Thought: Decomposing Physics for Truly Plausible Video Generation
Chain of Event-Centric Causal Thought for Physically Plausible Video Generation
The paper introduces an event-centric framework for Physically Plausible Video Generation (PPVG) that decomposes complex physical phenomena into causally linked event sequences. By integrating physics-driven reasoning with transition-aware cross-modal prompts, it achieves SOTA performance on PhyGenBench and VideoPhy benchmarks using off-the-shelf diffusion models like CogVideoX.
TL;DR
Generating a video of a glass ball sinking or a candle melting isn't just about appearance; it’s about causality. This paper introduces an event-centric framework that treats physical phenomena as a "Chain-of-Thought" for video. By breaking down a prompt into discrete events governed by real-world physics formulas and using "Interactive Keyframes" as visual anchors, the authors bridge the gap between photorealism and physical reality.
The Problem: Why Current AI "Fails" High School Physics
Modern video diffusion models like Sora or Kling produce stunning visuals but often stumble on the "Why" and "How" of motion.
- Causal Ambiguity: A prompt like "ice melting" is a process, not a state. Models often fail to capture the progression (solid -> slush -> liquid).
- Semantic Vagueness: Words are insufficient to describe the exact trajectory of a parabolic throw or the refraction index of glass. Without quantitative grounding, the "physics" in the video is just a hallucination.
Methodology: Solving Physics Through Decomposition
The authors argue that a video should be generated as a sequence of causally connected events. Their solution is two-fold:
1. Physics-driven Event Chain Reasoning (PECR)
Instead of feeding a raw prompt to the video model, the framework first performs Physics Formula Grounding. If the prompt involves a falling object, the system retrieves Newton’s laws to determine parameters. It then builds a Dynamic Scene Graph that tracks how objects' attributes (e.g., color, state, position) should change over time.
2. Transition-aware Cross-modal Prompting (TCP)
To ensure the transition between "Ice Cube" and "Water Puddle" is smooth, the model generates:
- Semantic Prompts: A narrative that evolves as the event progresses.
- Visual Prompts (The Secret Sauce): Instead of starting from random noise, the model creates Interactive Keyframes. It uses an image editor (like Qwen-Image-Edit) to "drag" or "modify" the scene based on physical formulas, providing the diffusion model with a strong visual prior for the denoising process.
Experiments: Proving the Physicality
The team tested their method on PhyGenBench and VideoPhy, benchmarks specifically designed to catch AI in physical lies.
- Quantitative Boost: They achieved a top score of 0.66 PCA (Physical Commonsense Alignment), significantly outperforming baseline models like CogVideoX and dedicated physics-aware models like DiffPhy.
- Ablation Insight: The most striking discovery was that Interactive Keyframe Synthesis (IKS) accounted for a massive chunk of the performance. Without these visual anchors, the model's physical alignment dropped by 17%.
Visual comparison: Notice how the "Honey Pouring" case (bottom row in Fig 5) shows a realistic rise in liquid level, a feat usually missed by standard models.
Critical Analysis & Future Outlook
While the "Chain of Event" approach is a massive leap forward, it isn't perfect. The paper identifies a key limitation: Compositional Physics. When multiple laws collide (e.g., a Newton’s cradle hitting a water balloon), the underlying LLMs still struggle to reason through the combined complexity.
Takeaway: The future of video generation isn't just "bigger models" or "more data." It's about inductive bias—integrating the mathematical laws of our universe into the generative process. This event-centric approach provides a scalable blueprint for building "World Simulators" that actually respect the rules of the world.