MemoBench: Visual Memory Evaluation Clips for World Models

MemoBench is a dataset for evaluating visual memory in world generation models. It contains 196 synthetic clips and 164 real clips depicting disappearance-and-reappearance scenarios. The dataset was created by tonyc54 and last updated on June 19, 2026.

Use Cases

• Benchmarking visual memory performance in AI models based on disappearance-and-reappearance scenarios
• Evaluating object permanence in generated video sequences
• Training models to track objects that leave and re-enter the camera view

Strengths

• Contains 196 synthetic clips and 164 real clips, providing a mix of data sources
• Includes metadata for exit and re-appear moments to facilitate structured evaluation
• Provides extracted frames and auxiliary files for evaluation

Limitations

• Row count and dataset size are unknown, which may limit suitability assessment
• Column-level documentation is absent; field semantics must be inferred after download

Provenance

Source

huggingface

Collection Method

Collection includes both synthetic and real video clips.

Freshness

Last updated 2026-06-19 03:58:55