Krea releases Krea 2 technical report

By sharing the end-to-end recipe for Krea 2, Krea proves that open-weights image models can move beyond generic "AI aesthetics" through modular training and thoughtful data curation. The report shows that RL and careful post-training alignment are just as critical for text-to-image generation as they are for LLMs.

• –**Steerability over defaults**: Instead of optimizing solely for a polished default aesthetic, Krea 2 focuses on creative exploration using a prompt expander (trained via RL with a diversity reward) and a style-reference system to minimize content leakage.
• –**Architectural efficiency**: The model streamlines the standard Diffusion Transformer (DiT) by replacing heavy per-block modulation MLPs with simple block-level biases (saving 20-30% of parameter count) and adopting GQA with gated sigmoid attention for stable training.
• –**Advanced post-training stack**: The training pipeline employs a custom DPO variant (STPO) to prevent policy divergence and high-frequency artifacts, followed by a multi-reward GRPO-style RL stage using rubric-based prompt evaluation and a dedicated structural artifact classifier.
• –**Custom database and training infra**: Krea built a custom PG-sharded metadata queue called "krablets" to scale data processing, and utilized FSDP2/Megatron-LM TP alongside the Weka filesystem to support aggressive checkpointing (completed in under 30 seconds).