Krea 2 Technical Report (www.krea.ai)

The Krea 2 Technical Report introduces a new series of foundation models aimed at enhancing creative exploration in image generation, overcoming the limitations of current AI systems that often converge on default aesthetics. By leveraging scalable transformer architectures and a comprehensive data infrastructure, Krea 2 focuses on producing expressive models that enable users to navigate a wide range of styles and visual compositions. This is achieved through a multi-stage training pipeline that includes innovative techniques such as improved variational autoencoders and reinforcement learning, allowing for enhanced output quality and user control. Significantly, Krea 2 implements a prompt expander and a style-reference system, which help bridge the gap between user intent and model output by allowing for richer visual directions and enhanced style mixing. The model is trained on a diverse and carefully curated dataset that intentionally excludes AI-generated images to prevent biases and improve alignment with user expectations. Krea 2 currently ranks among the top models on the Artificial Analysis leaderboard for text-to-image tasks, facilitating a more exploratory creative process for artists and content creators while maintaining competitive performance levels.