tl;dr: Krea released Krea 2, a 12B open-weights diffusion transformer for text-to-image generation that ranks in the top 10 on Artificial Analysis's leaderboard and 2nd among independent labs. The model uses a multi-stage pipeline (pretraining, midtraining, SFT, preference optimization, and RL with multiple reward models) on a curated dataset that explicitly excludes AI-generated images, paired with Qwen3-VL as text encoder, GQA with gated sigmoid attention, and a custom STPO variant of DPO. It ships with a prompt expander and style-reference system aimed at creative exploration rather than a single default aesthetic.