Research map / Video / Video-language understanding
Flash-VStream: Efficient Real-Time Understanding for Long Video Streams
Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Xiaojie Jin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Flash Memory (CSM + DAM) instead of Qwen2-VL under the same 11520 token budget | Qwen2-VL-7B · EgoSchema | accuracy (%): 64.0 -> 68.2 | Improved |
| Use Flash-VStream on Video-MME long videos with subtitles versus Qwen2-VL-online | Qwen2-VL-7B · Video-MME (w/ subs), Long | accuracy (%): 57.9 -> 61.4 | Improved |
| Use Flash-VStream (11520 tokens) versus Qwen2-VL full (24576 tokens) on MLVU | Qwen2-VL-7B · MLVU dev | accuracy (%): 66.0 -> 66.3 | Improved |
| Replace uniform sampling with K-means clustering memory (CSM) (ablation) | Flash Memory (CSM only) · EgoSchema | accuracy (%): 63.4 -> 66.8 | Improved |
| Add feature-centric DAM instead of uniform-sampled DAM (ablation) | Flash Memory · avg of MVBench, EgoSchema, Video-MME | average accuracy (%): 64.2 -> 64.9 | Improved |
| Add CSM clustering memory instead of uniform sampling (ablation) | Flash Memory · avg of MVBench, EgoSchema, Video-MME | average accuracy (%): 61.6 -> 63.6 | Improved |
| Use Euclidean distance instead of cosine similarity in DAM retrieval (ablation) | DAM · EgoSchema | accuracy (%): 66.5 -> 68.2 | Improved |
| Select top-k largest clusters instead of top-k smallest for DAM (ablation) | DAM · Video-MME (w/o subs) | accuracy (%): 59.6 -> 61.2 | Improved |
| Feature-centric retrieval instead of temporal-centric retrieval (ablation) | DAM · MVBench | accuracy (%): 64.4 -> 65.4 | Improved |
| Use DBScan instead of K-means for CSM (ablation) | CSM · EgoSchema | accuracy (%): 66.8 -> 66.6 | No clear effect |
| Increase DAM from 30 to 60 key frames (11520 to 19200 tokens) (ablation) | Flash Memory · avg of EgoSchema, MVBench, Video-MME | average accuracy (%): 64.9 -> 65.1 | No clear effect |
| Reduce CSM size from 60 to 15 clusters (ablation) | Flash Memory · avg of EgoSchema, MVBench, Video-MME | average accuracy (%): 63.6 -> 61.5 | Hurt |
Later papers that compared with this method
12 later accepted papers report Flash-VStream as a baseline in their results tables; 12 of them report a better number on most of the results they share with it.
- Question-guided Visual Compression with Memory Feedback for Long-Term Video Understanding (CVPR 2026 · better on 4 of 4 shared results)
- Progressive Online Video Understanding with Evidence-Aligned Timing and Transparent Decisions (ICLR 2026 · better on 3 of 3 shared results)
- RIVER: Real-time Video Interaction Benchmark (ICLR 2026 · better on 3 of 3 shared results)
- Streaming Video Crime Anticipation with Spatio-Temporal Causal Reasoning (CVPR 2026 · better on 3 of 3 shared results)
- StreamReady: Learning What to Answer and When in Long Streaming Videos (CVPR 2026 · better on 3 of 3 shared results)
- StreamKV: Streaming Video Question-Answering with Segment-based KV Cache Retrieval and Compression (AAAI 2026 · better on 1 of 1 shared result)
- CogStream: Context-guided Streaming Video Question Answering (AAAI 2026 · better on 1 of 1 shared result)
- Thinking in Streaming Video (ECCV 2026 · better on 1 of 1 shared result)
4 more not listed.
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Streaming Long Video Understanding with Large Language Models (NeurIPS 2024 · similarity 0.92)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.90)
- StreamForest: Efficient Online Video Understanding with Persistent Event Memory (NeurIPS 2025 · similarity 0.90)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.90)
- TimeChat: A Time-sensitive Multimodal Large Language Model for Long Video Understanding (CVPR 2024 · similarity 0.90)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
