Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai

Krea trained Krea 2 on **2–10 billion images**, using basic hashes before semantic deduplication and roughly **30–40 in-house filters**. Its pipeline progressed from **256px to 1K** before mid-training, supervised tuning, preference optimization, and reinforcement learning.

For image systems, treat dataset composition as a product decision. Audit aesthetic filters for style collapse, distill expensive vision-model judgments into cheap classifiers, and keep the training stack simple enough to test changes quickly.

These are research notes from one model rather than controlled comparisons of every lever. More reliable outputs may still trade away diversity, and the talk does not quantify Krea 2’s gains against alternatives.