CG-World: Large-Scale Dataset for World Models from Industrial CG Pipelines
CG-World has been launched by researchers as a comprehensive dataset and protocol that originates from industrial computer graphics production processes. In contrast to current datasets in video, robotics, and simulations that only partially represent states, actions, events, and observations, CG-World meticulously documents intermediate states, incorporating multimodal semantics, spatial configurations, skeletal and controller states, motion trajectories, camera and lighting settings, physics caches, contact events, and multi-pass renderings. The initial release, CG-World v1, features around 850,000 temporally synchronized segments lasting between 1 to 5 seconds. This dataset categorizes latent states, observations, relations, events, and branch metadata into cohesive spatiotemporal samples, aiming to enhance research in world models through its detailed, structured data from professional CG workflows.
Key facts
- CG-World is a large-scale world-state dataset and protocol from industrial computer graphics production pipelines.
- It explicitly records intermediate states including multimodal semantics, spatial structure, skeletal and controller states, motion curves, camera and lighting parameters, physics caches, contact events, and multi-pass renderings.
- CG-World v1 contains approximately 850,000 temporally aligned segments of 1-5 seconds.
- The dataset separates latent states, observations, relations, events, and branch metadata into unified spatiotemporal samples.
- It defines a branch lineage for intervention learning and counterfactual reasoning.
- Existing video, robotics, and simulation datasets capture only part of the joint dynamics of states, actions, events, and observations.
- The dataset is derived from industrial computer graphics production pipelines.
- CG-World supports intervention learning and counterfactual reasoning.
Entities
Institutions
- arXiv