The Case for Diversity Over Consistency

Production-grade image models often suffer from "mode collapse" to achieve consistency, resulting in bland, predictable outputs. Krea 2 intentionally trades this extreme reliability for stylistic range. By optimizing for faster generation, the model allows creative studios to explore visual concepts rather than being forced into the "average" aesthetic common in models like DALL-E 3 or Midjourney.

Data Curation as the Primary Lever

Once the architecture (typically a latent diffusion model) is locked, data quality becomes the sole differentiator. The Krea team employs a rigorous, multi-stage filtering pipeline to ensure the model learns robust concepts rather than artifacts:

  • Refusal of Synthetic Data: The team avoids training on AI-generated images to prevent "aesthetic stickiness" and the inheritance of biases from other models.
  • Automated Filtering: They utilize a combination of hash-based deduplication for scale and embedding-based methods (SSCD/CLIP) for near-duplicate removal.
  • Distilled Classifiers: Large Vision-Language Models (VLMs) are used to generate high-quality judgments, which are then distilled into lightweight, efficient classifiers capable of sweeping billions of images.
  • Sparse Autoencoders (SAEs): SAEs serve as an unsupervised tagging system, allowing the team to identify and filter out undesirable features like watermarks, signatures, and border artifacts.
  • World Knowledge: To ensure broad conceptual coverage, they rank Wikipedia articles by PageRank and use these concepts to guide data collection, ensuring the model understands important real-world entities.

The Training Pipeline

Krea 2 follows an LLM-inspired training progression:

  1. Resolution Scaling: Training begins at low resolution (256px) to learn semantic concepts before scaling up to 1k resolution for structural detail.
  2. Molding and Preference Optimization: After pre-training, the model undergoes supervised fine-tuning (SFT) on curated datasets (photography, graphic design, etc.) followed by preference optimization to align the model with specific aesthetic goals.
  3. Reinforcement Learning: Similar to RLHF in LLMs, they use reward servers to teach the model better anatomy and text rendering.
  4. Prompt Expansion: A small, autoregressive language model is trained to expand short user prompts into detailed, descriptive ones, which are more "in-distribution" for the diffusion model, leading to higher-quality outputs.

Future Directions

As vision-language models improve, the team is moving toward more structured conditioning. By leveraging better VLM capabilities, they are exploring ways to condition models on bounding boxes and scene graphs, moving beyond simple text prompts to give users more granular control over image composition.