Context As Memory: Scene-Consistent Long Video Generation Data

Developed by KlingTeam for SIGGRAPH Asia 2025, this dataset supports scene-consistent interactive long video generation through memory retrieval. It is distributed as a multi-part archive containing video sequences designed for maintaining visual coherence over extended durations.

Use Cases

• Training generative models for long-form video synthesis using memory retrieval features
• Evaluating scene consistency in interactive video generation tasks
• Benchmarking temporal coherence in AI-generated video sequences

Strengths

• Peer-reviewed origin from SIGGRAPH Asia 2025
• Specifically formatted for memory-retrieval-based video synthesis
• Supports research into long-duration temporal coherence

Limitations

• Requires manual file merging via command line before use
• Specific metadata schema and column names are not explicitly documented in the summary
• Potential high storage and compute requirements for long-form video processing

Provenance

Source

KlingTeam, SIGGRAPH Asia 2025

Freshness

Last updated October 2025.