Krea 2's feature set is designed around a core philosophy: creative exploration over deterministic output. The model prioritizes breadth and visual surprise alongside quality and control. This is an intentional design decision, not a limitation. Every feature reflects this: diversity in output, style reference as the primary control, moodboards for open-ended direction, and a training set built from real images.

Aesthetic Diversity by Design

Most image models converge on a default interpretation of any given prompt. Krea 2 was built to diverge. Give it "horse rearing" and you get four distinct compositions: different styles, camera angles, lighting moods, and artistic treatments. This is not randomness. The training pipeline (pretraining, midtraining, SFT, preference optimization, reinforcement learning) was tuned to balance breadth with quality at every stage.

Style Reference System

Upload a single reference image and the model generates output that captures the aesthetic (color palette, texture, mood, composition style) without copying the content. This is the primary creative control mechanism. It replaces long style descriptions in text prompts with visual examples. The model deeply understands aesthetics because that was the explicit training objective.

Moodboard Generation

Not everything can be conveyed with a single reference. Curate a collection of images and the model synthesizes themes, motifs, and styles across the full set into new output. A moodboard combining "Cyber Zine" textures, "Minimalist Ink" linework, and "Thermal Airbrush" color grading produces images that blend all three directions. This maps directly to how creative teams already work.

Zero Synthetic Training Data

The pretraining set contains no AI-generated or synthetic images. This is unusual. Most frontier models include synthetic data to fill gaps or boost specific capabilities. Krea chose not to because synthetic data can flatten aesthetic range and create homogeneous output. The result is output that community members have called "the most uncensored nonlobotomized open source image model" available.

Two Checkpoints: Raw + Turbo

Krea 2 Raw is the undistilled mid-training checkpoint. It has a malleable latent space that accepts LoRA fine-tuning with genuine room to push the model toward new domains. Krea 2 Turbo is distilled to 8 inference steps and generates 2K images in about 2 seconds. Use Raw for maximum quality and training flexibility. Use Turbo when you need speed for rapid iteration or bulk production.

GRPO-Trained Prompt Expander

The built-in prompt expander was trained using Group Relative Policy Optimization, the same technique used for reasoning models. It enriches short prompts with visual detail before generation. This is why a three-word prompt like "samurai drawing blade" can produce a fully realized cinematic composition with dramatic lighting, wide-angle perspective, and atmospheric depth.