i1

i1 is a 3B-parameter text-to-image diffusion model from Princeton (zlab), presented as a simple and fully open recipe for training strong text-to-image models. It was trained only on publicly available datasets, and the project releases checkpoints, training and inference code, and the data processing pipeline.

The accompanying study runs 300+ controlled experiments (over 700K TPU v6e hours) on modeling and data choices. Findings include that equal weighting is a strong default for mixing curated datasets, and that larger text encoder adapters improve results with few extra parameters. The baseline uses a LightningDiT-style backbone with QK-norm and long skip connections, a T5Gemma-2B text encoder, the FLUX.2 VAE, and long synthetic captions generated with Qwen3-VL-30B-A3B.

The authors report that i1 is competitive with leading models on GenEval, DPG, PRISM, CVTG-2K and LongText, and beats the best previous fully open model by 29.5 absolute points on average. A YouTube reviewer noted it trails models such as Z-Image and Qwen-Image in quality, so its main value is as a research foundation for studying text-to-image training from scratch rather than as a top-quality production model.