Research & Papers

i1: Fully open 3B-parameter text-to-image model beats closed rivals by 29.5%

Open-source diffusion model trained only on public data outperforms top closed-source alternatives.

Deep Dive

A team of researchers (Boya Zeng et al.) has released i1, a 3B-parameter text-to-image diffusion model that is fully open—including weights, training data, and code. The model is trained exclusively on publicly available datasets, a stark contrast to most state-of-the-art models that rely on proprietary data or undisclosed recipes. The team conducted a systematic investigation with over 300 controlled experiments, consuming more than 700,000 TPU v6e hours, to identify optimal modeling and data design choices. Key findings include that equal weighting is a strong default for mixing curated datasets, and that larger text encoder adapters improve performance with minimal added parameters.

i1 achieves impressive results across five representative benchmarks (GenEval, DPG, PRISM, CVTG-2K, and LongText), outperforming the best existing fully open model by an average of 29.5 absolute percentage points. While not explicitly compared to top closed models like DALL-E 3 or Midjourney, its performance is described as “competitive.” The project provides a complete open recipe: model checkpoints, training and inference code, and the data processing pipeline. This transparency establishes a practical foundation for future open research in text-to-image generation, enabling reproducibility and community-driven improvements.

Key Points
  • i1 is a 3B-parameter diffusion model trained only on public datasets, fully open-sourced.
  • Based on 300+ controlled experiments (700K+ TPU hours) to find optimal configurations.
  • Outperforms best existing fully open model by 29.5 absolute percentage points on average across 5 benchmarks.

Why It Matters

Democratizes state-of-the-art image generation with a fully open recipe, accelerating research and reproducibility.

📬 Get the top 10 AI stories daily