Krea publishes the Krea 2 technical report, a creative exploration model.

Krea 2, an open-source image model by Krea detailed in Sangwu Lee's technical report, ranks in the top 10 on Artificial Analysis using a multi-stage pipeline.

With the open-sourcing of Krea 2's weights, Krea is publishing the model's technical report, authored by Sangwu Lee. The guiding principle: to make image generation a medium for exploration. The report observes that as image models have gained in reliability, many have converged towards a narrow default aesthetic, efficient for production but lacking when it comes to seeking a style, a mood, or a composition. Krea 2 therefore aims for expressiveness—covering a wide range of aesthetics—and controllability—the ability to navigate them.

To achieve this, the report describes three components: a base model with broad stylistic coverage; a style reference system, which allows specifying a rendering from one or more images rather than solely through words, with strength adjustment and style mixing, and minimal content leakage; and a prompt expander, which enriches a short prompt with more detailed visual direction without overriding the user's intent.

Regarding methodology, Krea claims a multi-stage pipeline modeled after LLM training, from pretraining to RL, and a clear-cut choice regarding data: no AI-generated images in the pre-training mix, deemed to introduce bias and cap quality. The report places Krea 2 in the top 10 of Artificial Analysis's text-to-image ranking, and second among independent labs. It accompanies the two open-weights checkpoints already available, Raw and Turbo.