Princeton's i1: A fully open text-to-image model backed by 300 controlled experiments
机器之心 · wechat · 2026-08-23
Princeton University's Zhuang Liu team released i1, a 3B parameter text-to-image model. Backed by 300+ controlled experiments and 700k TPU hours, it uses fully open code and data. i1 outperforms the best previous fully open models by an average of 29.5% on five benchmarks, with excellent text rendering capabilities. Key findings include: large adapters on a single encoder can match multi-encoder setups; long caption training requires prompt rewriters for short prompts; high-res training doesn't need broad data coverage to preserve capabilities.
More from Models
- Stefano Ermon teases 'largest and most capable model' landing next week, opens early access — StefanoErmon · 2026-10-16
- Inception teases its biggest model yet after Mercury 2 makes TIME's Best Inventions — StefanoErmon · 2026-10-16
- Diffusion LLM Mercury 2 named to TIME's Best Inventions of 2026 — StefanoErmon · 2026-10-16
- Gemini 4 Argon scores 97.5% on ARC-AGI-3, second lab past 95% milestone — fchollet · 2026-10-16
- New inference engine 'e' hits 10,500 tok/s on a single DGX Station — brandon_galang · 2026-10-16
- NVIDIA's Nemotron 3.5 Super VL hits CoreWeave: 12B-active multimodal MoE at $0.25/M input tokens — wandb · 2026-10-16