README.md

Paper: Can GPTs Evaluate Graphic Design Based on Design Principles?

Daichi Haraguchi¹, Naoto Inoue¹, Wataru Shimoda¹, Hayato Mitani², Seiichi Uchida², Kota Yamaguchi^1
1CyberAgent.Inc, ² Kyushu University
Accepted to SIGGRAPH Asia (Technical Communications Track). [Project-page]

Introduction

This repository contains the data for "Can GPTs Evaluate Graphic Design Based on Design Principles?".

Data

./GPT_eval_data
├── gpt_eval
      ├── gpt_abs_alignment.csv
      ...
      └── gpt_comp_whitespace.csv
├── human_eval_data
      ├── human_abs_alignment.csv
      ...
      └── human_comp_whitespace.csv
└── images.zip

Each CSV file includes the evaluation results. File names consist of "[eval method]_[eval type]_[design principle].csv."
Note:
"eval. method" is "gpt" or "human."
"eval. type" is "abs" (absolute evaluation) or "comp" (comparative (relative) evaluation).
"design principle" is "alignment," "overlap," or "whitespace."

Example:
"gpt_abs_alignment.csv" is the absolute evaluation result by GPT for alignment.

"images.zip" includes images.
The structure in the "images.zip" is below.

images
├── lefttop_large
      ├── [image_ID].png
      ...
      └── [image_ID].png
...
└── org
      ├── [image_ID].png
      ...
      └── [image_ID].png

Note:
The directory name consists of "[perturbation element]_[perturbation size]"
"Org" includes original images without perturbation.
Each image name is an image ID that is written in the CSV files.

The structure of each CSV file

CSV for absolute evaluation

id	perturbation	0	1	2	3	4	avg
image ID	perturbation size	score	score	score	score	score	average score

CSV for relative evaluation

id	comparative	better_design_0	better_design_1	better_design_2	better_design_3	better_design_4	voting_res
image ID	perturbation size of comparison image	voting result	voting result	voting result	voting result	voting result	aggregation of voting results

Visualization

We provides a notebook for showing results.

• scatter_plot.ipynb shows the scatter plot between human evaluation and GPT evaluation.

Reference

@misc{
coming soon
}

Contact

This repository is maintained by Daichi Haraguchi.
E-mail: haraguchi_daichi_xa[at]cyberagent.co.jp