LLaVA-NeXT Video Captioning Tool

• (LLaVA-Video-7B-Qwen2)

1. Environment Setup

• Clone this repository and navigate to the LLaVA folder

  git clone https://github.com/LLaVA-VL/LLaVA-NeXT
  cd LLaVA-NeXT

• Create and activate the Conda environment

  conda create -n llava python=3.10 -y
  conda activate llava

• Install the inference package

  pip install --upgrade pip
  pip install -e ".[train]"

2. Run Captioning

• Script: llava_captioning.py

  python llava_captioning.py

3. Output

• Outputs are saved as a JSONL file (1 line = 1 video).
• It reads the existing output and skips already-processed video_name entries to resume (resume).

Notes

• llava_captioning.py assumes the following structure:
  • /home/dataset/video_eval/L{1..5}/{short,medium,long}/*.mp4