i1

i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models

Official code for i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models

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][model][dataset][project page]

We investigate the design space of text-to-image diffusion models to understand how modeling and data choices affect model capabilities. This exploration culminates in i1, a 3B-parameter model that performs competitively with leading models at 1024-resolution, as measured by the average percentage score across GenEval, DPG-Bench, PRISM, CVTG-2K, and LongText-Bench.

Showcase

Open-Source Plan

We fully open-source the training code, data, and recipes for reproducing our i1-3B model.

• 3B Model Checkpoint (1024-Resolution) [PyTorch] [JAX]
• 1B Model Checkpoint (1024-Resolution) [PyTorch] [JAX]
• Intermediate 256- and 512-Resolution Checkpoints [3B] [1B]
• JAX Training and Inference Code for TPUs and GPUs
• PyTorch Training and Inference Code for GPUs
• Dataset (Captions and Image-Caption TFRecords) and Data Pipelines
• Benchmark Evaluation Pipeline
• Multi-Aspect-Ratio Checkpoint, Data Pipelines, and Training Code

Quick Start

Install PyTorch inference environment

conda create -n i1_torch_infer python=3.11 -y
conda activate i1_torch_infer
python -m pip install torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
python -m pip install numpy==1.26.4 pillow tqdm transformers==4.57.1 diffusers==0.35.1 accelerate safetensors sentencepiece

Generate image with your custom prompt

git clone https://github.com/zlab-princeton/i1
cd torch_inference
python generate.py \
    --prompt "Render the following text at the center of the image on a clean background: 'Flow on, river! flow with the flood-tide, and ebb with the ebb-tide! Frolic on, crested and scallop-edg'd waves!'"

Code Structure

This codebase contains five independent folders.

• benchmark_eval contains the instructions and code for running evaluations with the 5 benchmarks we use.
• data_processing contains the code for downloading images, recaptioning images, and creating TFRecord files for the image-caption pairs.
• jax contains the training and inference code for our controlled experiments and the final i1-3B model in JAX.
• torch_inference contains the inference code for the final i1-3B model in PyTorch.
• torch_train contains the training code for the final i1-3B model in PyTorch.

Acknowledgement

We gratefully thank the Google TPU Research Cloud (TRC) program for providing the primary computing resources for this project. Additional support was provided by the Princeton Research Computing resources at Princeton University, which are managed by a consortium of groups led by the Princeton Institute for Computational Science and Engineering (PICSciE) and Research Computing. We would like to thank Liang-Chieh Chen, Ishan Misra, Kaiming He, Yida Yin, Haozhe Chen, Wenhao Chai, Linrong Cai, Linzhan Mou, and Xingyu Fu for valuable discussions and feedback. We also thank Yufeng Xu, Shengbang Tong, Yiyang Lu, and Hanhong Zhao for helpful discussion on TPU. We are grateful to Cihang Xie's research group for sharing their JAX DiT codebase, which served as the launching point for our research. This repository is built using the big_vision, transformers, and diffusers codebases.

Citation

If you find this repository helpful, please consider citing:

@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}
}