README.md

# VGI-Bench | Data

[![arXiv](https://img.shields.io/badge/arXiv-2608.19583-b31b1b?style=for-the-badge&logo=arxiv&logoColor=white)](https://arxiv.org/abs/2608.19583)  [![Project Page](https://img.shields.io/badge/Project-Page-1f8acb?style=for-the-badge&logo=googlechrome&logoColor=white)](https://hexuan21.github.io/VGI-Bench/)

This repository holds the **data** of VGI-Bench, in two subtrees:

- `bench_data/` -- the benchmark itself: 27 tasks, each at 3
  difficulty levels (`lv1`/`lv2`/`lv3`) with 10 instances for each level, where
  an instance pairs an input image (16:9 -- also the first frame for video
  models) with a text prompt. 16 of the tasks additionally ship an
  image-generation prompt, the rest being video-only, and every task comes with
  the rubric and completeness specs used for the VLM-as-judge evaluation.
- `eval_res/` -- the raw generations every evaluated model produced on those
  instances, plus the 640px previews the project page streams.

```

bench_data/<task>/

  input_image/lv{1,2,3}/<task>_lv<N>_<id>.png   input image = first frame

  prompt/video_gen_prompt/                       prompt for video-gen models

  prompt/image_gen_prompt/                       prompt for image-gen models / UMM (optional)

  judge_spec/_judge_rubric_<task>.json           VLM-as-judge checklist

  judge_spec/_completeness_judge.txt             completeness scale (0 / 1 / 2)



eval_res/

  video_model_res/<model>/<task>/<task>_<lv>_<id>.mp4

  image_model_res/<model>/<task>/<task>_<lv>_<id>.png

  _preview/...                                   640px previews

```