Real-time Physics in 3D Scenes with Language
Published:Dec 31, 2025 17:32
•1 min read
•ArXiv
Analysis
This paper introduces PhysTalk, a novel framework that enables real-time, physics-based 4D animation of 3D Gaussian Splatting (3DGS) scenes using natural language prompts. It addresses the limitations of existing visual simulation pipelines by offering an interactive and efficient solution that bypasses time-consuming mesh extraction and offline optimization. The use of a Large Language Model (LLM) to generate executable code for direct manipulation of 3DGS parameters is a key innovation, allowing for open-vocabulary visual effects generation. The framework's train-free and computationally lightweight nature makes it accessible and shifts the paradigm from offline rendering to interactive dialogue.
Key Takeaways
- •Enables real-time, physics-based 4D animation of 3D scenes.
- •Uses a Large Language Model (LLM) to translate language prompts into executable code.
- •Directly manipulates 3D Gaussian Splatting (3DGS) parameters.
- •Avoids time-consuming mesh extraction and offline optimization.
- •Train-free and computationally lightweight, making it accessible.
Reference
“PhysTalk is the first framework to couple 3DGS directly with a physics simulator without relying on time consuming mesh extraction.”