Krea releases Krea 2 technical report and open-weights models
The independent lab behind Krea 2 has published a detailed technical report and released open-weights for its K2 Raw and K2 Turbo text-to-image models, which now rank in the top 10 on the Artificial Analysis leaderboard.
Krea has published a comprehensive technical report detailing Krea 2 (K2), a series of open-weights text-to-image foundation models designed to address the narrow aesthetic defaults prevalent in previous image generation systems. The release includes K2 Raw and K2 Turbo model weights under a permissive license, offering researchers and developers access to a system built for broad aesthetic diversity and user control.
The Krea 2 architecture is built on a diffusion transformer (DiT) framework that incorporates grouped-query attention, sigmoid-gated attention, and lightweight timestep modulation. To enhance text encoding and latent space representation, the larger models utilise Qwen 3 VL as the final text encoder and the FLUX 2 VAE. This design aims to provide expressive capabilities while maintaining training stability and efficiency, moving away from the complex stacks of interdependent models that often characterise previous diffusion systems.
Training the models required a rigorous multi-stage pipeline involving pretraining, midtraining, supervised fine-tuning, preference optimisation, and reinforcement learning. A key innovation in this process is a novel preference optimisation variant called STPO, developed to mitigate policy divergence. The reinforcement learning stage employs a multi-reward GRPO-style method, utilising specific reward models for aesthetics, prompt-following, text-rendering, and artifact detection to ensure high-fidelity outputs.
To support this scale of research, Krea developed a custom data warehousing system called 'krablet' based on PostgreSQL to manage training datasets. The company also switched from Ceph to the Weka filesystem to improve stability and checkpointing speed at scale. These infrastructure changes were critical for managing the fault tolerance and hardware stability challenges inherent in large-scale distributed training.
Krea 2 ranks in the top 10 on the Artificial Analysis leaderboard for text-to-image models and places second among independent labs. The release includes a prompt expander and a style-reference system to enhance user control, allowing creators to navigate a broad visual space rather than receiving a single polished default. The models are designed to serve as a comprehensive baseline for creative generative experiences while maintaining competitive performance.
