Authors: Hyungjin Chung, Hyelin Nam, Jiyeon Kim, Hyojun Go, Byeongjun Park, Junho Kim, Joonseok Lee, Seongsu Ha, and Byung-Hoon Kim
This repo contains the official implementation of the paper "Video Parallel Scaling: Aggregating Diverse Frame Subsets for VideoLLMs".
You can set up the environment using
pip install -r requirements.txt
Note that we used CUDA 12.8 in all our experiments.
./examples contains a demo video entire_003.mp4 from the EventHallusion benchmark. Our demo runs VPS on this example video. If you wish to try out other videos in Video-MME or EventHallusion, prepare the data accordingly.
You can simply run
./scripts/demo.sh
to run the demo. We offer three different model classes
model=
gemma-3-{}b-it
Qwen2.5-VL-{}B-Instruct
InternVL3-{}B
You may choose from existing model sizes that are uploaded on huggingface. You may also change the number of frames to be sampled from the raw video.
The following command will run the baseline VideoLLM
# Baseline
python demo.py \
--model_name ${model} \
--video_path ${video_path} \
--num_frames ${num_frames} \
--prompt "Summarize the video in one sentence."
whereas the following will run VPS with $J = $ num_parallel_inputs - 1
# VPS
python demo.py \
--model_name ${model} \
--video_path ${video_path} \
--num_frames ${num_frames} \
--prompt "Summarize the video in one sentence." \
--use_vps \
--num_parallel_inputs 2
If you find our work interesting, please consider citing
@article{chung2025video,
title={Video Parallel Scaling: Aggregating Diverse Frame Subsets for VideoLLMs},
author={Chung, Hyungjin and Nam, Hyelin and Kim, Jiyeon and Go, Hyojun and Park, Byeongjun and Kim, Junho and Lee, Joonseok and Ha, Seongsu and Kim, Byung-Hoon},
year={2025},
journal={arXiv preprint arXiv:2509.08016},
}