README.md

metadata

task_categories:
  - video-text-to-text
language:
  - en
tags:
  - video-understanding
  - streaming-video
  - real-time
  - long-video
  - vision-language-model

StreamingVLM Datasets

This repository contains datasets used in the paper StreamingVLM: Real-Time Understanding for Infinite Video Streams.

StreamingVLM introduces a model designed for real-time, stable understanding of infinite visual input. It addresses challenges like escalating latency and memory usage in processing long video streams by maintaining a compact KV cache and aligning training with streaming inference. The project includes novel datasets for both training and evaluation, particularly Inf-Streams-Eval, a new benchmark with videos averaging over two hours that requires dense, per-second alignment between frames and text.

• Paper: https://huggingface.co/papers/2510.09608
• Code: https://github.com/mit-han-lab/streaming-vlm
• Project/Demo Page: https://streamingvlm.hanlab.ai

Included Datasets

The datasets associated with the StreamingVLM project include:

• Inf-Stream-Train: This dataset is used for supervised fine-tuning (SFT) of the StreamingVLM model.
• Live-WhisperX-526K: An additional dataset utilized during the SFT process, described as Livecc_sft in the project's setup.
• Inf-Stream-Eval: A new benchmark dataset introduced for evaluating real-time video understanding, featuring long videos (averaging over two hours) and requiring dense, per-second alignment between frames and text.

Sample Usage: Dataset Preparation for SFT

To prepare the necessary datasets for Supervised Fine-Tuning (SFT) as described in the GitHub repository:

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

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}, 
}