README.md

---
license: apache-2.0
task_categories:
- text-to-video
language:
- en
tags:
- benchmark
- design
- multimodal
- graphic-design
- animation
size_categories:
- n<1K
configs:
  - config_name: reference
    data_files:
      - split: full_layout
        path: all_full_layout/renders/**
      - split: single_components
        path: all_single_components/renders/**
  - config_name: sora2
    data_files:
      - split: full_layout
        path: generated_videos/sora2/full_layout/**
      - split: single_components
        path: generated_videos/sora2/single_components/**
  - config_name: veo31
    data_files:
      - split: full_layout
        path: generated_videos/veo3.1/full_layout/**
      - split: single_components
        path: generated_videos/veo3.1/single_components/**
---
# GDBench-Video: Evaluating Design Video Generation

Benchmark for evaluating video generation models on graphic design videos. Built on 136 real animated design layouts from the [Lica dataset](https://lica.world/). 

**Paper: [arXiv:2605.16223](https://arxiv.org/pdf/2605.16223) | Code: [github.com/purvanshi/lica-bench](https://github.com/purvanshi/lica-bench/tree/main/scripts/motion_gen_eval)**


## Usage
```
from datasets import load_dataset

ds = load_dataset("lica-world/GDB-video")
```

## Citation
```
@article{deganutti2026evaluating,
  title={Evaluating Design Video Generation: Metrics for Compositional Fidelity},
  author={Deganutti, Adrienne and Cao, Dingning and Seol, Jaejung and Hirsch, Elad and Mehta, Purvanshi},
  journal={arXiv preprint arXiv:2605.16223},
  year={2026}
}
```