Datasets:

lica-world
/

GDB

license: apache-2.0
task_categories:
  - visual-question-answering
  - image-to-text
  - text-to-image
language:
  - en
tags:
  - benchmark
  - design
  - multimodal
  - graphic-design
  - svg
  - typography
  - layout
  - animation
  - lottie
pretty_name: 'GDB: GraphicDesignBench'
size_categories:
  - 1K<n<10K
configs:
  - config_name: template-4
    data_files:
      - split: train
        path: template-4/train-*
  - config_name: template-5
    data_files:
      - split: train
        path: template-5/train-*

GDB: GraphicDesignBench

39 benchmarks for evaluating vision-language models on graphic design tasks — layout, typography, SVG, template matching, animation. Built on 1,148 real design layouts from the Lica dataset.

Paper: arXiv:2604.04192 | Code: github.com/lica-world/GDB | Blog: lica.world

Usage

from datasets import load_dataset

ds = load_dataset("lica-world/GDB", "svg-1")

Schema

Field	Type	Description
`sample_id`	string	Sample identifier
`benchmark_id`	string	e.g. `svg-1`, `typography-3`
`domain`	string	layout, typography, svg, template, temporal, category, lottie
`task_type`	string	`understanding` or `generation`
`prompt`	string	Evaluation prompt
`ground_truth`	string	Expected answer (JSON for complex types)
`image`	Image	Input image (when applicable)
`metadata`	string	Task-specific fields as JSON

Evaluation

pip install git+https://github.com/lica-world/GDB.git

from gdb.registry import BenchmarkRegistry

registry = BenchmarkRegistry()
registry.discover()
bench = registry.get("svg-1")
scores = bench.evaluate(predictions, ground_truth)

Citation

@article{gdb2026,
  title={GDB: A Real-World Benchmark for Graphic Design},
  author={Deganutti, Adrienne and Hirsch, Elad and Zhu, Haonan and Seol, Jaejung and Mehta, Purvanshi},
  journal={arXiv preprint arXiv:2604.04192},
  year={2026}
}

Apache 2.0