← back
Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai
Takeaway
Generative image quality and stylistic range depend heavily on deliberate data curation and accurate captions.
Summary
- Krea 2 prioritizes fast visual exploration and stylistic diversity, contrasting this goal with highly consistent but visually homogeneous image generation.
- Latent diffusion compresses images before transformer processing to reduce the cost of modeling full-resolution pixels.
- Data curation removes duplicates and rebalances overrepresented concepts while preserving unconventional aesthetics that standard quality filters might discard.
- The captioning pipeline combines OCR and optional metadata with detailed vision-language captions; the team avoids synthetic images that can imprint other models' aesthetics.
diffusiondata-curationkrea
Original description
The most reliable way to render a person is to render the most boring average person and put them in the center of the frame. Sangwu Lee offers that as the price the big image models pay for consistency: ask a production model for a burning skull and every output comes back clean, competent, and nearly identical. Krea 2, whose medium variant is now open source, trades the other way, optimizing for fast generation and stylistic range so that a studio that does not yet know what it wants can actually explore. Most of the talk is about data, which he says twice over is basically everything once the architecture is locked. The examples are specific. A painting photographed on a wall is perfectly good training data except that captioners consistently omit the frame and the white wall behind it, so the model learns to hang every painting it generates. They refuse to train on AI generated images at all, because the aesthetic is sticky and you inherit somebody else's model. Deduplication runs on hashes first across two to ten billion images, then on embeddings for near duplicates. A large vision language model's judgment gets distilled down into a classifier cheap enough to sweep a billion images. Sparse autoencoders double as an unsupervised tagging system for catching watermarks and border artifacts. World knowledge coverage is checked against Wikipedia concepts ranked by PageRank. Thirty to forty in house filters in total. Speaker info: https://github.com/RE-N-Y https://re-n-y.github.io/devlog/ https://github.com/krea-ai/krea-2 Timestamps: 0:00 - Open sourcing Krea 2 medium 1:40 - Consistency versus diversity in production models 3:23 - How diffusion models train, and why latent space 5:59 - Data is basically everything 6:53 - Bad data, and why they refuse AI images 8:34 - The captioning pipeline, and the painting on a white wall 10:15 - Deduplication and cheap classifiers at billion image scale 11:56 - Sparse autoencoders as an unsupervised tagging system 13:39 - Wikipedia PageRank for world knowledge coverage 14:35 - The training pipeline, borrowed wholesale from LLMs 18:54 - What actually mattered for iterating fast 19:46 - The stack is inverting back toward DALL-E 2