How Krea’s Small Team Trains Its Own Image Models | Victor Perez

In an interview with Krea co-founder and CEO Victor Perez, the discussion centers on how the small independent company achieved a top-four result on the Arena leaderboard for its Krea 3 Ultra model using a team of only six or seven researchers. Perez attributes this success to a sharp focus on image models, rigorous data curation for training and editing pairs, and significantly greater compute efficiency than larger labs. He also highlights the importance of art direction and human taste in diffusion model evaluations compared to objective metrics used in large language models. Open article →