Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai
Aug 18, 2026 · 21:46
Sangwu Lee of Krea.ai explains what went into training Krea 2, its open-sourced image foundation model, arguing that once architecture is locked, data is everything. To keep stylistic diversity over the 'most boring average person' consistency of production models, Krea filters billions of images: no AI images, OCR and vision-language captions, hash and embedding dedup, distilled VLM classifiers, sparse autoencoders as unsupervised taggers for watermarks and borders. World knowledge is checked against Wikipedia concepts by PageRank. Training goes from 256 to 1K resolution through pre-training, mid-training, SFT, preference optimization, and GRPO-style RL, plus a prompt expander. Lee wants a single clean transformer without VAEs/text encoders and VLM-generated bounding boxes or scene graphs.
