Krea 2 training deep-dive: data is everything, and never train on AI images

AI Engineer · youtube · 2026-08-18

A talk by Krea co-founder Sangwu Lee on the tradeoffs and data engineering behind Krea 2 (medium variant now open source).

Core thesis

Production image models pay for consistency with boring output: the most reliable person render is the most average person centered in frame. Krea 2 optimizes the other way—fast generation and stylistic range so studios that don't yet know what they want can explore.

Data engineering (the bulk of it)

Once architecture is locked, "data is basically everything."
Pitfall example: paintings photographed on walls are fine training data, but captioners consistently omit the frame and white wall, so the model learns to hang frames on every painting it generates.
They refuse to train on AI-generated images at all—the aesthetic is sticky and you inherit somebody else's model.
Dedup runs on hashes first across 2–10 billion images, then on embeddings for near duplicates.
A large VLM's judgment gets distilled 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; 30–40 in-house filters in total.
The training pipeline is borrowed wholesale from LLMs; the stack is inverting back toward the DALL-E 2 era.

More from Multimodal