MemoBench is a dataset for evaluating visual memory in world generation models. It focuses on disappearance-and-reappearance scenarios, where a target object is visible, leaves the camera view, and later reappears.
The dataset currently contains:
196 synthetic clips
164 real clips
Metadata for exit and re-appear moments
Extracted frames and auxiliary files for evaluation
Synthetic Examples
Real… See the full description on the dataset page: https://huggingface.co/datasets/tonyc54/MemoBench.