Can a small academic team build a strong text-to-image model using only public datasets? Introducing i1: a simple, fully open recipe for strong text-to-image models

We study the design space of text-to-image diffusion models w/ 300+ controlled experiments. The insights are combined to build i1-3B, which outperforms the best prior fully open model by 29.5 absolute percentage points on average across 5 benchmarks.

We release the checkpoint, training/inference code, dataset, data pipeline, and full recipe. This is not just another checkpoint. It is a strong baseline people can study, reproduce, modify, and build on.

At 1024px, i1-3B is competitive with leading open-weight T2I models across our benchmark suite, and outperforms models like FLUX.1 [Dev] and HiDream-I1-Full. This is not just on older benchmarks (GenEval and DPG-Bench). It also holds on newer benchmarks (PRISM, CVTG-2K, and LongText-Bench) used by leading models (e.g., Z-Image, Ideogram 4.0) and better reflect modern T2I capabilities.

Alongside strong general image generation, i1-3B is the first fully open T2I model with competitive text rendering performance (to the best of our knowledge). With 3B parameters, it scores 92% on LongText-Bench. For comparison, the best prior fully open model scores 15%.

The strong performance of i1-3B comes from 300+ controlled experiments carefully ablating modeling and data designs. We highlight some of our findings below:

Modeling findings: 1) a single strong text encoder + larger text encoder adapter > combining multiple text encoders 2) AdaLN provides marginal benefit for T2I relative to its large number of parameters 3) dual-stream DiT + long skip connections is a strong backbone design

Data findings: 1) train on long captions + inference-time prompt rewrite > train on short captions, even for short inference prompts 2) equal weighting is a strong default for mixing curated datasets 3) with a diverse mix of datasets, fewer images and more epochs don’t hurt performance much 4) strong high-resolution performance doesn't require broad coverage in high-resolution training data

i1 shows that strong text-to-image models do not have to rely on opaque recipes, inaccessible data, or highly specialized new components. A simple, fully open recipe built from public datasets and carefully validated design choices can be competitive with leading open-weight models. Find many more results and analyses in our paper! Led by @zeng_boya with many others contributing, Tianze Luo, Jucheng Shen, Shu Pu, @TaimingLu and @GabrielSarch

Check i1 out: Model: Data: Project Page: Code: Paper:

@XingyuFu2 Nice work and write-up, congrats!

@XingyuFu2 Thanks Lucas!

great work!

Great work!

Nice results and ablations! Congrats!

Super cool

Definitely the trend seems the same. As it was for CNNs and vision. First very deep models, then the refined ones with only imp layers that can run on edge devices.

@jon_barron Do the team plan to release 1b model soon?

Great work for democratizing T2I model training! Will definitely take a look.

Nice work Zhuang! Congratulations!

Thanks Xudong!

@ricklamers 😎