Dragonjinny/FiFA-pickapic-v2

Pick-a-Pic v2 · FiFA Filtered Subsets

These subsets were produced by filtering the original Pick-a-Pic v2 dataset using FiFA, a data filtering algorithm proposed in the paper Automated Filtering of Human Feedback Data for Aligning Text-to-Image Diffusion Models.

Overview

The filtering process is based on three key metrics:

1. Preference Margin: Estimated using PickScore
2. Text Quality: Estimated through LLM scoring
3. Text Diversity: Estimated using K-NN distance

Dataset Configurations

This dataset provides several configurations, each corresponding to a different number of filtered triplets selected by the automated FiFA algorithm. You can choose from FiFA-500, FiFA-1k, FiFA-5k, FiFA-10k, FiFA-20k, FiFA-50k, and FiFA-100k, depending on your needs. The only difference between these configurations is the number of examples included; all are filtered using the same FiFA method. We recommend using the FiFA-5k configuration as the default, as it generally works best for Stable Diffusion training.

Quick Start

from datasets import load_dataset

# Load a specific configuration
dataset = load_dataset("Dragonjinny/FiFA-pickapic-v2", "FiFA-5k", split="train")

# Access the data
import io
from PIL import Image

for example in dataset:
    caption = example["caption"]  # The prompt text
    jpg_0 = example["jpg_0"]      # First image (bytes)
    jpg_1 = example["jpg_1"]      # Second image (bytes)
    label_0 = example["label_0"]  # Binary label (0 or 1) indicating which image is preferred
    
    # Convert bytes to PIL Images
    image1 = Image.open(io.BytesIO(jpg_0)).convert("RGB")
    image2 = Image.open(io.BytesIO(jpg_1)).convert("RGB")
    
    # Now you can work with the images
    print(f"Caption: {caption}")
    print(f"Preferred image: {'jpg_0' if label_0 == 1 else 'jpg_1'}")
    # image1.show()  # Display the first image
    # image2.show()  # Display the second image

Data Format

Each example contains:

• caption: The prompt text
• jpg_0: First image (bytes)
• jpg_1: Second image (bytes)
• label_0: Binary label (0 or 1) indicating which image is preferred
• __index_level_0__ : Unique ID for each data point

Citation

If you use this dataset in your research, please cite our paper:

@inproceedings{
yang2025automated,
title={Automated Filtering of Human Feedback Data for Aligning Text-to-Image Diffusion Models},
author={Yongjin Yang and Sihyeon Kim and Hojung Jung and Sangmin Bae and SangMook Kim and Se-Young Yun and Kimin Lee},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=8jvVNPHtVJ}
}

License

This dataset is licensed under MIT License, following the license of the original Pick-a-Pic v2 dataset.