Princeton drops i1, a fully open 3B image generation model

Princeton researchers have rolled out i1, a 3B-parameter text-to-image diffusion model. At around 12GB, the entire model fits comfortably on standard consumer hardware. It handles photorealistic scenes, diverse artistic and anime styles, and shows surprisingly solid capabilities when rendering intricate text inside images.

What it does and how it stacks up

Let’s get one thing straight: i1 is not claiming the text-to-image crown. The Princeton team is refreshingly upfront that i1 trails leading systems like Z Image and Qwen Image, and likely sits behind recent releases like Ideogram and ERNIE Image. If your only goal is generating pretty pictures for a project, this probably won't replace your daily driver.

Where it shines is in the realm of truly open research.

Across five representative benchmarks—GenEval, DPG, PRISM, CVTG-2K, and LongText—i1 holds its ground against top-tier architectures. More impressively, it outperforms the best existing fully open model by an average of 29.5 absolute percentage points.

A complete, transparent recipe

Most modern "open" releases are open in weights only. They ship pre-trained checkpoints while keeping data curation pipelines and precise hyperparameters closely guarded secrets. Outside researchers are left guessing how the models were actually built.

Princeton took the opposite route by putting everything on the table.

Along with the model checkpoints and inference code, the researchers released full training code and the complete data processing pipeline, relying exclusively on publicly available datasets.

To figure out what actually works, the team ran more than 300 controlled ablation experiments totaling over 700,000 TPU v6e hours. They walked away with actionable design rules: for example, equal weighting across curated datasets serves as an exceptionally robust default, and using larger text encoder adapters substantially boosts generation quality with negligible parameter overhead.

Anyone looking to spin it up can grab the checkpoints, pipeline code, and setup instructions straight from the project repository linked in the paper.

Why it matters

If you just need an AI canvas for everyday art, i1 isn't meant for you. There are sharper, larger models available right now.

However, for university labs, independent developers, and engineering teams planning to train an image model from the ground up, i1 is an invaluable blueprint. It lifts the veil on text-to-image data pipelines and training choices that tech giants usually treat as trade secrets.

We think Princeton is doing the broader research ecosystem a massive favor here. By demonstrating that publicly available data and rigorous ablations can still yield a competitive modern model, i1 establishes an honest, reproducible baseline that the open community can finally build on.