Flash-VStream
Flash-VStream: Efficient Real-Time Understanding for Long Video Streams
Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Xiaojie Jin · 2025-06-30
- Management
- T2 Dynamic Memory
- Memory changes during deployment through a stable management policy.
- Representation
- Source- and modality-specific
How experience becomes memory
Each path is one memory component, traced from what is stored to how it is found again.
Context Synopsis Memory
- Represented asModality-specific encodings & geometric stateSource- and modality-specific
- Stored inHierarchical multi-tier store
- Retrieved bySimilarity search, Latent / parametric activation
- Source evidenceSource not retained
A fixed-capacity context memory stores low-resolution visual-feature cluster centroids and temporal positions.Full paper: 3.1. Model Architecture · PDF p. 3
Detail Augmentation Memory
- Represented asModality-specific encodings & geometric stateSource- and modality-specific
- Stored inHierarchical multi-tier store
- Retrieved bySimilarity search, Latent / parametric activation
- Source evidenceSource not retained
High-resolution visual features nearest to high-density context clusters are retained for detailed spatial augmentation and jointly activated in the language model.Full paper: A. Implementation Details · PDF p. 12
How memory changes (T2 operations)
Update · Rule-based update
A fixed K-means update incorporates each incoming low-resolution frame feature while maintaining a bounded centroid set. For Flash-VStream, this specifically supports fixed-rule insertion or replacement of Context Synopsis Memory.
Consolidation · Merging & deduplication
Semantically similar frames are consolidated into cluster centroids to remove temporal redundancy. For Flash-VStream, this specifically supports merging or deduplication of overlapping entries of Context Synopsis Memory.
Related systems
- DigimonGPTT2 · Video & streaming
- DriveWAM Selective KV MemoryT2 · Video & streaming
- FilmAgentT2 · Video & streaming
- KubrickT2 · Video & streaming
- Latent Spatial MemoryT2 · Video & streaming
- LUMA-RAGT2 · Video & streaming