StreamingVLM: Real-Time Understanding for Infinite Video Streams

Paper | Slides | Demo Page

🧠 TL;DR

StreamingVLM enables real-time, stable understanding of effectively infinite video by keeping a compact KV cache and aligning training with streaming inference. It avoids quadratic cost and sliding-window pitfalls, runs up to 8 FPS on a single H100, and wins 66.18% vs GPT-4o mini on a new long-video benchmark. It also boosts general VQA without task-specific finetuning. You can grasp the gist by skimming this section first.

🎬 Demo

Go to streamingvlm.hanlab.ai to see more cases and try our model.

https://github.com/user-attachments/assets/1a15b496-55c5-4c66-809d-a49d70e5d864

🛠️ Install

./scripts/env_infer.sh
./scripts/env_sft.sh

You can set up the environment by running the scripts above.

🚀 Inference

You can run inference by the command below.

conda activate streamingvlm-infer
python streaming_vlm/inference/inference.py

📚 SFT

Prepare Dataset

First, download mit-han-lab/Inf-Stream-Train to /path/to/your/Inf-Stream-Train. Then, download chenjoya/Live-WhisperX-526K to /path/to/your/Inf-Stream-Train/Livecc_sft. Preprocess the LiveCC dataset with the following command:

cd $DATASET_PATH/Livecc_sft
find . -type f -exec mv -t . {} +

Download mit-han-lab/Inf-Stream-Eval to /path/to/your/Inf-Stream-Eval.

Finally, set environment paths:

export DATASET_PATH=/path/to/your/Inf-Stream-Train
export EVAL_DATASET_PATH=/path/to/your/Inf-Stream-Eval

You can prepare data by following the steps in order.

▶️ Run SFT

*You can kick off SFT by executing the scripts below.*

conda activate streamingvlm-sft
./scripts/sft_stage_1.sh
./scripts/sft_stage_2.sh # High Quality Annealing Data

📊 Evaluation

Efficiency

conda activate streamingvlm-infer
./scripts/eval_efficiency.sh  

You can benchmark efficiency by running the script above.

OVOBench

First, make the OVOBench data structure like:

data/ovobench
├── AutoEvalMetaData
├── COIN
├── cross_task
├── Ego4D
├── hirest
├── MovieNet
├── OpenEQA
├── ovo_bench_new.json
├── perception_test
├── star
├── thumos
├── youcook2
└── YouTube_Games

Then, prepare the OVOBench environment and run evaluation:

./scripts/env_ovo.sh
conda activate streamingvlm-ovo
./scripts/eval_OVOBench.sh

You can start OVOBench eval by these commands.

VQA

We use VLMEvalKit to evaluate VQA tasks.

conda activate streamingvlm-infer
./scripts/eval_VQA.sh

You can launch VQA evaluation with the script above.

Inf-Stream-Eval

conda activate streamingvlm-infer
./scripts/eval_Inf-Stream-Eval.sh

You can run the in-house eval by calling this script.

LiveSports3k-cc

conda activate streamingvlm-infer
export LIVESPORTS3K_PATH=/path/to/your/LiveSports-3K/videos
conda activate streamingvlm-infer
./scripts/eval_LiveSports3k-cc.sh

You can evaluate LiveSports3k-cc with the path set above.

Modify FPS

If you would like to change inference FPS, use the following command:

sed -i 's/^FPS = .*/FPS = float(os.environ.get("QWENVL_FPS", "2.0"))/' \
  "$(python -c 'import inspect,qwen_vl_utils.vision_process as m; import os; print(os.path.abspath(inspect.getsourcefile(m)))')"

You can tweak FPS by editing the line via the command above.

Citation

If you find StreamingVLM useful or relevant to your project and research, please kindly cite our paper:

@misc{xu2025streamingvlmrealtimeunderstandinginfinite,
      title={StreamingVLM: Real-Time Understanding for Infinite Video Streams}, 
      author={Ruyi Xu and Guangxuan Xiao and Yukang Chen and Liuning He and Kelly Peng and Yao Lu and Song Han},
      year={2025},
      eprint={2510.09608},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.09608}, 
}