STRIDE

STRIDE (Structured Temporal Refinement with Iterative DEnoising) is a lightweight proactive activation model for streaming video understanding. It employs a masked diffusion module at the activation interface to jointly predict and progressively refine activation signals over a sliding temporal window, producing temporally coherent proactive responses in online streaming scenarios.

• Paper release
• Model weights release (STRIDE-2B)
• Demo website
• Training code
• Evaluation scripts

Download and extract the source videos from the interlive/stream-data dataset into a video_root directory:

uv run scripts/download_videos.py --output_dir /path/to/video_root

Pass --max_per_folder N to extract only N videos per folder for a quick sample (default 0 extracts everything).

Convert the source annotations into masked-diffusion activation training data (.jsonl):

uv run --no-sync scripts/prepare_activation_dataset.py \
    --video_root /path/to/video_root \
    --output_path /path/to/data.jsonl

--video_root points to a directory organized as video_root/source/video_id.

bash scripts/train.sh /path/to/data.jsonl

The first argument is the path to the activation data produced in the previous step.

Run ET-Bench (trigger detection + scoring):

ANNO_PATH=/path/to/etbench_txt_v1.0.json DATA_PATH=/path/to/videos \
    bash scripts/etbench.sh interlive/STRIDE-2B

ANNO_PATH is the ET-Bench annotation JSON and DATA_PATH is the video root.

For other benchmarks (OVO-Bench, StreamingBench), refer to lmms-eval.

@article{kim2026stride,
  title={STRIDE: When to Speak Meets Sequence Denoising for Streaming Video Understanding},
  author={Kim, Junho and Lee, Hosu and Rehg, James M. and Kim, Minsu and Ro, Yong Man},
  journal={arXiv preprint arXiv:2603.27593},
  year={2026}
}

This project is released under the Apache 2.0 License.