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

Model Overview

i1 is a 3-billion-parameter open-source text-to-image diffusion model developed by researchers at ZLab, Princeton University. Introduced in June 2026 (arXiv:2606.11289), i1 addresses the lack of transparency in modern generative vision models by providing a completely open recipe: open model weights (1B and 3B variants), full training and inference code in PyTorch and JAX, data processing pipelines, and a massive 160M+ dataset of synthetically annotated image-caption pairs (i1-captions).

Developed through more than 300 systematic ablation experiments across 700,000+ TPU v6e compute hours, i1 demonstrates how principled choices in model architecture, text adapters, and dataset curation can achieve performance competitive with top-tier open-weight models at 1024x1024 resolution.

Key Technical Features & Architecture

• Diffusion Transformer Backbone: Employs an XL/2-sized LightningDiT cross-attention backbone with Query-Key (QK) normalization and long skip connections for high-fidelity spatial generation.
• Text Conditioning Adapter: Integrates a T5-Gemma-2B text encoder, leveraging a larger adapter design to significantly boost prompt understanding and fine-grained text adherence.
• Visual VAE: Incorporates the FLUX.2 VAE for high-density latent encoding and decoding of 1024x1024 image resolution.
• Synthetic Caption Pipeline: Utilizes i1-captions, curated from 12 open datasets and re-captioned with detailed synthetic descriptions using Qwen3-VL-30B-A3B.
• Dual PyTorch & JAX Implementations: Includes full source code for training and inference across both JAX (TPU-optimized) and PyTorch (GPU-optimized) frameworks.

Performance & Benchmarks

i1 substantially outpaces prior fully open-source baselines and achieves competitive alignment and visual quality against leading open-weight models across multiple standard benchmarks:

• GenEval: Evaluates compositional fidelity, object attributes, and spatial relationships.
• DPG-Bench: Measures compliance with complex multi-sentence prompts.
• PRISM: Assesses general visual quality, alignment, and aesthetic appeal.
• CVTG-2K: Tests in-image visual text generation performance.
• LongText-Bench: Evaluates adherence to extended, detailed descriptions.

Quickstart & Usage

1. Environment Setup

conda create -n i1_infer python=3.11 -y
conda activate i1_infer

# Install dependencies
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

2. PyTorch Inference CLI

git clone https://github.com/zlab-princeton/i1
cd i1/torch_inference

# Generate image from prompt
python generate.py \
  --prompt "A serene mountain lake at sunrise, highly detailed, realistic oil painting style"

Official Links & Resources

• Project Page / Website: zlab-princeton.github.io/i1
• arXiv Preprint: arXiv:2606.11289
• Paper PDF: Download PDF
• GitHub Repository: zlab-princeton/i1
• Hugging Face Model Checkpoints: zlab-princeton/i1-3B
• Hugging Face Dataset: zlab-princeton/i1-captions