OpenWorldLib: A Unified Codebase and Definition of Advanced World Models

OpenWorldLib is a standardized inference framework for advanced world models, addressing the lack of a unified definition in the field.

The paper proposes that a world model is a model or framework centered on perception, equipped with interaction and long-term memory capabilities, for understanding and predicting the complex world.

It categorizes essential capabilities and integrates tasks like interactive video generation, 3D generation, multimodal reasoning, and vision-language-action (VLA) into a single framework.

The framework includes modules for Operator, Synthesis, Reasoning, Representation, and Memory, with a Pipeline for unified execution.

Evaluations on NVIDIA A800 and H200 GPUs show that while methods like Hunyuan-WorldPlay excel in interactive video generation, challenges remain in physical consistency and 3D geometric stability.

The paper also clarifies that tasks like text-to-video generation are not considered world models due to lack of complex perceptual input.

Future directions include improving efficiency of next-frame prediction and leveraging LLMs as foundational backbones for world models.

OpenWorldLib aims to standardize research and facilitate fair comparisons in world model development.

Source code and documentation are publicly available.