DEsignBench Prompts contains the prompt table released with DEsignBench: Exploring and Benchmarking DALL-E 3 for Imagining Visual Design. It provides visual design text-to-image prompts with original user inputs, expanded prompts, and requested aspect ratios.
This Hugging Face loader covers the prompt TSV only. It does not include the DEsignBench gallery images or generated model outputs.
The dataset is intended for text-to-image prompt evaluation in visual design scenarios. Models can use userinput or expandprompt as generation prompts and evaluate generated images against the design intent described by each row.
No active public leaderboard is bundled with this Hugging Face dataset.
The prompts are primarily written in English, with some prompts or expanded prompt text containing multilingual rendered text examples.
{
"id": "text_0_0",
"demo_imgname": "text_0_0",
"userinput": "a graffiti art of the text \"free the pink\" on a wall",
"expandprompt": "Photo of a smooth stone wall with the graffiti art 'free the pink' painted in a gradient from pink to blue. Surrounding the text are intricate patterns and a silhouette of a city skyline at the base.",
"aspectratio": "wide"
}
id (string): Stable row identifier, copied from demo_imgname.demo_imgname (string): Original DEsignBench demo image name.userinput (string): Original user prompt.expandprompt (string): Expanded text-to-image prompt.aspectratio (string): Requested aspect ratio. The source values are wide, square, and tall.The dataset exposes one benchmark split:
| split | rows |
|---|---|
| test | 215 |
The source aspect-ratio counts are:
| aspectratio | rows |
|---|---|
| wide | 173 |
| square | 36 |
| tall | 6 |
The prompt TSV is distributed by the DEsignBench project page. DEsignBench was created as a text-to-image benchmark for visual design scenarios and includes categories for design technical capability and design application scenarios.
This loader exposes the original prompt metadata from the released TSV. It does not add new annotations.
The dataset contains natural-language prompts for generated visual design tasks. It may include names of places, objects, events, brands, or text that a generated image should render. The dataset card does not identify personal information in the released prompt table.
The prompt set reflects the benchmark authors' visual design categories and prompt-writing process. Models evaluated with these prompts may overfit to the represented prompt distribution, aspect ratios, and design scenarios.
The gallery images released separately by the original project follow the terms of the image generation models that produced them. This Hugging Face dataset contains prompts only.
The source Hugging Face dataset and project release do not declare a prompt dataset license, so this dataset is marked as unknown.
@article{lin2023designbench,
title={DEsignBench: Exploring and Benchmarking DALL-E 3 for Imagining Visual Design},
author={Kevin Lin and Zhengyuan Yang and Linjie Li and Jianfeng Wang and Lijuan Wang},
journal={arXiv preprint arXiv:2310.15144},
year={2023}
}
Thanks to Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, and Lijuan Wang for creating DEsignBench.
4 commits
DEsignBench Prompts contains the prompt table released with DEsignBench: Exploring and Benchmarking DALL-E 3 for Imagining Visual Design. It provides visual design text-to-image prompts with original user inputs, expanded prompts, and requested aspect ratios.
This Hugging Face loader covers the prompt TSV only. It does not include the DEsignBench gallery images or generated model outputs.
The dataset is intended for text-to-image prompt evaluation in visual design scenarios. Models can use userinput or expandprompt as generation prompts and evaluate generated images against the design intent described by each row.
No active public leaderboard is bundled with this Hugging Face dataset.
The prompts are primarily written in English, with some prompts or expanded prompt text containing multilingual rendered text examples.
{
"id": "text_0_0",
"demo_imgname": "text_0_0",
"userinput": "a graffiti art of the text \"free the pink\" on a wall",
"expandprompt": "Photo of a smooth stone wall with the graffiti art 'free the pink' painted in a gradient from pink to blue. Surrounding the text are intricate patterns and a silhouette of a city skyline at the base.",
"aspectratio": "wide"
}
id (string): Stable row identifier, copied from demo_imgname.demo_imgname (string): Original DEsignBench demo image name.userinput (string): Original user prompt.expandprompt (string): Expanded text-to-image prompt.aspectratio (string): Requested aspect ratio. The source values are wide, square, and tall.The dataset exposes one benchmark split:
| split | rows |
|---|---|
| test | 215 |
The source aspect-ratio counts are:
| aspectratio | rows |
|---|---|
| wide | 173 |
| square | 36 |
| tall | 6 |
The prompt TSV is distributed by the DEsignBench project page. DEsignBench was created as a text-to-image benchmark for visual design scenarios and includes categories for design technical capability and design application scenarios.
This loader exposes the original prompt metadata from the released TSV. It does not add new annotations.
The dataset contains natural-language prompts for generated visual design tasks. It may include names of places, objects, events, brands, or text that a generated image should render. The dataset card does not identify personal information in the released prompt table.
The prompt set reflects the benchmark authors' visual design categories and prompt-writing process. Models evaluated with these prompts may overfit to the represented prompt distribution, aspect ratios, and design scenarios.
The gallery images released separately by the original project follow the terms of the image generation models that produced them. This Hugging Face dataset contains prompts only.
The source Hugging Face dataset and project release do not declare a prompt dataset license, so this dataset is marked as unknown.
@article{lin2023designbench,
title={DEsignBench: Exploring and Benchmarking DALL-E 3 for Imagining Visual Design},
author={Kevin Lin and Zhengyuan Yang and Linjie Li and Jianfeng Wang and Lijuan Wang},
journal={arXiv preprint arXiv:2310.15144},
year={2023}
}
Thanks to Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, and Lijuan Wang for creating DEsignBench.
4 commits