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

Training high-performance image foundation models like Krea 2 requires prioritizing data curation over architecture, as data quality remains the primary determinant of model output. Unlike production-grade models that often suffer from mode collapse to ensure reliability, Krea 2 focuses on stylistic diversity and faster generation to support creative exploration. The training pipeline utilizes latent diffusion models, incorporating sophisticated filtering techniques such as hash-based deduplication, visual language model-driven captioning, and sparse autoencoders for unsupervised tagging to remove undesirable artifacts. Beyond pre-training, the model undergoes supervised fine-tuning, preference optimization, and reinforcement learning to enhance anatomical accuracy and text rendering. Future advancements in image generation will likely shift toward more complex conditioning methods, such as bounding boxes and scene graphs, leveraging the increasing capabilities of visual language models to provide more precise control over generated content.