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]

1. Overview

This dataset contains all captions used in our controlled experiments and the final training of the i1 model. Detailed instructions for downloading the corresponding images and matching the image-caption pairs can be found in our codebase.

2. Dataset Info

We use 12 curated image datasets and generate synthetic captions for text-to-image training.

2.1 Subsets

2.2 Columns

(1) caption1, caption2, caption3, caption4, caption5:
Long Qwen3-VL-30B-A3B captions used in our controlled experiments and the final training of the i1 model. At each training iteration, we randomly sample one caption for each image. Due to resource constraints, not all datasets have five sets of captions.
(2) qwen2vl_2b, qwen2.5vl_3b, qwen3vl_2b, qwen3vl_4b:
Captions generated by Qwen2-VL 2B, Qwen2.5-VL 3B, Qwen3-VL-2B, and Qwen3-VL-4B. They are used in our ablation of VLMs as synthetic captioners (see Section 5.1 of our paper).
(3) short:
Short Qwen3-VL-30B-A3B captions used in our ablation of prompt length (see Section 5.1 of our paper).
(4) no_center_crop:
Long Qwen3-VL-30B-A3B captions based on raw images without center-cropping or resizing. They are used in our ablation of image preprocessing before synthetic captioning (see Appendix D.1 of our paper).

3. Quick Start

Load a subset with:

from datasets import load_dataset
ds = load_dataset("zlab-princeton/i1-captions", "fluxreason", split="train")

4. Usage

To use the captions for model training with the i1 training codebase, please follow our data processing pipeline to combine the captions with their corresponding images into TFRecords.

Alternatively, we also provide processed TFRecords that can be directly used:

256×256 Resolution Datasets

512×512 Resolution Datasets

1024×1024 Resolution Datasets

The processed TFRecords for YFCC, RedCaps, and Megalith are currently not provided due to image license constraints. Please follow our data processing pipeline to create them.

5. 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}
}