Physical-ICL is a project-level dataset repository for studying physical in-context learning in video generation.
The current version contains a preliminary subset built from Physics-IQ. Each sample is organized as a query video with candidate demonstration videos. These demonstrations are labeled by their relationship to the query, such as good, weak, opposite, or irrelevant demonstrations.
