TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design
Python
11
3 commits
updated Jun 1, 2026
This is the code and data repository for the paper
TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design arXiv:2605.20731 · PDF
and the accompanying TASTE dataset, hosted on the Hugging Face Hub:
TASTE is a corpus of designer-panel rankings of AI-generated graphic designs across multiple aesthetic and description-faithfulness dimensions, plus a small preference model trained on it.
This is a joint work of Lica World and Contra.
The repo is organized into three top-level parts, each self-contained with its own README:
| Directory | What it is | README |
|---|---|---|
analysis/ | The dataset-paper analysis — signal-validation distribution tests, the open-weight VLM-as-judge benchmark, and hallucination-flag agreement. | analysis/README.md |
data/ | Processing scripts for the TASTE dataset on the Hugging Face Hub — download a local snapshot and derive the per-dimension ranking CSVs and pairwise battles used by analysis/ and taste-scorer/. | data/README.md |
taste-scorer/ | The pip-installable TASTE preference model — inference for per-dimension "which design does the panel prefer?" probabilities from an image pair and a prompt. | taste-scorer/README.md |
The three parts are independent; install only what you need.
# Score image pairs with the preference model
cd taste-scorer && pip install -e .
# Reproduce the paper analysis
cd analysis && pip install -r requirements.txt
# Fetch and process the dataset
cd data && pip install -r requirements.txt
The dataset itself is not committed here; it is fetched from the Hugging
Face Hub (see data/README.md). Designer identities are
masked before release — do not attempt to re-identify raters.
purvanshi/TASTEIf you use the TASTE dataset or the TASTE-scorer, please cite the paper (arXiv:2605.20731).
@article{zhu2026taste,
title={TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design},
author={Zhu, Haonan and Hirsch, Elad and Minetti, Alexandria and Nulty, Allison and Mehta, Purvanshi},
journal={arXiv preprint arXiv:2605.20731},
year={2026}
}
MIT — see LICENSE.
TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design
Python
11
3 commits
updated Jun 1, 2026
This is the code and data repository for the paper
TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design arXiv:2605.20731 · PDF
and the accompanying TASTE dataset, hosted on the Hugging Face Hub:
TASTE is a corpus of designer-panel rankings of AI-generated graphic designs across multiple aesthetic and description-faithfulness dimensions, plus a small preference model trained on it.
This is a joint work of Lica World and Contra.
The repo is organized into three top-level parts, each self-contained with its own README:
| Directory | What it is | README |
|---|---|---|
analysis/ | The dataset-paper analysis — signal-validation distribution tests, the open-weight VLM-as-judge benchmark, and hallucination-flag agreement. | analysis/README.md |
data/ | Processing scripts for the TASTE dataset on the Hugging Face Hub — download a local snapshot and derive the per-dimension ranking CSVs and pairwise battles used by analysis/ and taste-scorer/. | data/README.md |
taste-scorer/ | The pip-installable TASTE preference model — inference for per-dimension "which design does the panel prefer?" probabilities from an image pair and a prompt. | taste-scorer/README.md |
The three parts are independent; install only what you need.
# Score image pairs with the preference model
cd taste-scorer && pip install -e .
# Reproduce the paper analysis
cd analysis && pip install -r requirements.txt
# Fetch and process the dataset
cd data && pip install -r requirements.txt
The dataset itself is not committed here; it is fetched from the Hugging
Face Hub (see data/README.md). Designer identities are
masked before release — do not attempt to re-identify raters.
purvanshi/TASTEIf you use the TASTE dataset or the TASTE-scorer, please cite the paper (arXiv:2605.20731).
@article{zhu2026taste,
title={TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design},
author={Zhu, Haonan and Hirsch, Elad and Minetti, Alexandria and Nulty, Allison and Mehta, Purvanshi},
journal={arXiv preprint arXiv:2605.20731},
year={2026}
}
MIT — see LICENSE.