Krea 2 Technical Report

Krea 2 is a series of foundation models designed for creative image generation that prioritizes aesthetic diversity and user control, addressing the industry trend toward narrow default aesthetics. The models employ a multi-stage training pipeline including pretraining, supervised finetuning, and reinforcement learning, along with architectural innovations like diffusion transformers and improved text encoders. To bridge the gap between model capabilities and user intent, Krea 2 includes a prompt expander and style-reference system that allow creators to explore different visual directions through both text and image inputs.