Krea 2: Technical Report on the Open-Weight 12B Image Model

Krea has published a technical report on the Krea 2 series, designed to target diverse aesthetic styles and high user controllability. Beyond simply generating high-quality images, this model aims to serve as an exploratory medium that lets users explore a variety of styles and compositions.

Key Technical Features:

• Model Architecture: Built on a Diffusion Transformer (DiT) structure, incorporating state-of-the-art techniques such as iREPA, GQA (Grouped-Query Attention), and Sigmoid-gated Attention to improve training efficiency and stability.
• Multi-stage Training Pipeline: The model's output distribution was optimized through a sophisticated pipeline spanning Pretraining, Midtraining, SFT (Supervised Finetuning), Preference Optimization, and RL (Reinforcement Learning).
• User Control System: Offers a Prompt Expander that expands simple prompts into rich visual directions, and a Style-reference system that allows precise injection of an image's style and mood.

Krea 2 has entered the top 10 in the Text-to-Image category on the Artificial Analysis leaderboard, ranking 2nd among independent lab models, demonstrating its performance. The model weights are released under a Permissive license.