i1-datasets/i1-midjourneyv6-1024-resolution-1m-tfrecord

i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models
Boya Zeng, Tianze Luo, Shu Pu, Jucheng Shen, Taiming Lu, Gabriel Sarch, Zhuang Liu
Princeton University
[arXiv][code][model][project page]

Overview

To prepare the dataset for training, we store the image-caption pairs as TFRecords.

This HuggingFace dataset contains the TFRecords corresponding to the midjourneyv6 dataset at 1024×1024 resolution. Concretely, we only retain raw images with a shorter edge of at least 1024 pixels, center-crop them to squares, resize them to 1024×1024, and randomly subsample the resulting dataset to 1M images.

It also serves as an example of what a dataset processed using our data processing pipeline looks like.

You can find processed TFRecords for other datasets and resolutions here.

Structure

dataset-train.tfrecord-*-of-00128 are TFRecord shards, where 128 is the total number of shards.
dataset_info.json and features.json are TFDS metadata files needed by tfds.builder_from_directory(...) to load and decode the dataset correctly.

Download

pip install -U "huggingface_hub"

hf download i1-datasets/i1-midjourneyv6-1024-resolution-1m-tfrecord \
    --repo-type dataset \
    --local-dir /path/to/save/dataset

Citation

If this dataset is useful for your research, please cite the following work:

@article{zeng2026i1,
  title={i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models},
  author={Zeng, Boya and Luo, Tianze and Pu, Shu and Shen, Jucheng and Lu, Taiming and Sarch, Gabriel and Liu, Zhuang},
  journal={arXiv preprint arXiv:2606.11289},
  year={2026}
}