Princeton Team Releases Fully Open Text-to-Image Model i1, Backed by 300+ Controlled Experiments

Current text-to-image models are advancing rapidly, but most works independently introduce entire model, data, and training designs, using private data for training and lacking thorough ablation studies and reproducible recipes, making it difficult for researchers to determine whether superior performance stems from data or design.

The i1 model from Princeton University's Zhuang Liu team addresses this need by providing a fully open and reproducible model from training code to training data, offering researchers a unified starting point for controlled experiments.

The team conducted controlled experiments based on a unified baseline (shallow fusion, frozen text encoder, flow matching-trained DiT network, combining 12 public datasets with VLM-generated long captions), changing only one design at a time.

On the model side, they found: multiple text encoders can improve performance, but increasing a single adapter can achieve similar effects at lower cost; AdaLN has many parameters but limited additional gains; at the same parameter count, dual-stream outperforms single-stream, which outperforms cross-attention, and U-ViT long skip connections consistently improve performance.

On the data side, they found: short caption training weakens performance, long caption training underperforms on short prompts, which can be mitigated with a prompt rewriter; when data is diverse, small-scale repetition has little impact; high-resolution training does not require extensive data coverage and still retains text rendering ability.

Based on the conclusions from controlled experiments, the team combined simple designs to train the 3B-parameter i1, which averages 29.5 percentage points ahead of the previous best fully open model on five benchmarks: GenEval, DPG-Bench, PRISM-Bench, CVTG-2K, and LongText-Bench.

i1's overall performance is close to some larger, open-weight-only models, such as 20B Qwen-Image.

i1 is the first fully open model to achieve relatively accurate text rendering, scoring 0.8531 on CVTG-2K and 0.922 on LongText-Bench.

i1 demonstrates that training a competitive text-to-image model does not necessarily rely on private data or opaque recipes; using only public datasets and systematically validated designs can achieve a strong level of performance.

i1's reproducible recipe provides a unified starting point for text-to-image research, enabling researchers to evaluate new designs under controlled conditions.

However, i1 still approaches but does not surpass open-weight-only models, and the evaluation focuses on prompt following and text rendering, without fully covering general image quality.