Dataset Card for CyclePrefDB-T2I

Project page | Paper | Code

Cycle consistency-based preferences dataset consists of 468K comparison pairs for text-to-image generation. Input text prompts are from sDCI and images are generated using 4 different text-to-image diffusion models with 3 different random seeds. Reconstruction similarity is measured with SBERT (higher score means more similar).

We also have a dataset of cycle consistency preferences for image-to-text generation which you can find here

Loading CyclePrefDB-T2I

You can use the Hugging Face Datasets library to load the CyclePrefDB-T2I.

from datasets import load_dataset

# Load dataset
dataset = load_dataset("carolineec/CyclePrefDB-T2I")

Dataset Format

Each of the train, validation, and test splits in the dataset have the following format:

{
	"prompt": ## input text prompt (string)
	"generations": [image1, image2, ...] ## list of candidate PIL images (list of PIL Images)
	"reconstructions": ["recon 1", "recon 2", ...] ## list of reconstructed text descriptions paths corresponding with the image generations. (list of strings)
	"sbert_score": [0.5, 0.6, ...] ## list of float SBERT scores between the corresponding reconstructions and input text prompt. Higher scores indicates higher similarity. (list of floats)
}

Citation

@article{bahng2025cyclereward,
title={Cycle Consistency as Reward: Learning Image-Text Alignment without Human Preferences},
author= {Bahng, Hyojin and Chan, Caroline and Durand, Fredo and Isola, Phillip},
journal={arXiv preprint arXiv:2506.02095},
year={2025}
}