What is Krea2? The Krea 2 image model family explained.

The short answer

Krea 2 is a series of text-to-image foundation models from Krea AI. Its stated goal is wide aesthetic diversity plus practical creative control: users can steer what appears through text and steer how it looks through prompt expansion, creativity levels, style references, and moodboards. Krea says the family was trained from scratch rather than wrapping another image generator.

Think of Krea 2 as two related product lines: hosted creative models for using Krea’s controls, and open checkpoints for local inference, research, and LoRA work.

Krea2 model comparison: Medium vs Large vs RAW vs Turbo

When should you choose Krea 2 Large instead of Medium?

Start with Krea 2 Medium when you are still deciding the subject, crop, palette, or style. It is the practical hosted default for iteration and expressive illustration. Move a selected direction to Krea 2 Large when the final image depends on finer photographic texture, difficult materials, convincing skin or fabric, raw grain, motion blur, or low-dynamic-range character.

Large is not automatically the better choice for every prompt. If composition or art direction is still unresolved, higher fidelity only makes the wrong decision more expensive. Keep the prompt, aspect ratio, seed, and creativity level stable when comparing Medium and Large so the model change is the variable you are actually judging.

What is inside the open Krea2 image model?

Krea’s technical report describes a 12-billion-parameter dense diffusion-transformer backbone. The open release uses a Qwen Image VAE and Qwen3-VL text encoder with multi-layer feature aggregation. The report also documents grouped-query attention, sigmoid-gated attention, a single-stream transformer design, multi-stage pretraining and post-training, reinforcement learning, and timestep distillation.

Those architecture details explain what the model is; they do not automatically describe every hosted feature. Prompt expansion and the style-reference system are additional steering systems built around the base model. Krea also says editing, stronger image reference, and native 2K/4K generation remain areas of future work, so this page does not present Krea 2 as a universal precision editor.

Why prompts and references behave differently

A prompt communicates content and explicit visual constraints. Krea’s prompt expander can map a short user request into a denser model-facing caption while trying to preserve intent. A style reference communicates a visual language that can be hard to name—palette, surface, medium, lighting, and compositional habits. The technical report says the style system supports multiple references, continuous strength, weighted mixing, and reduced content leakage.

Who should choose Krea 2?

Choose it when exploration and art direction matter: campaign concepts, editorial frames, illustration systems, fashion studies, architecture materials, moodboards, and LoRA-led custom styles. Choose an edit-specific or identity-preserving model when exact text, fixed product geometry, a specific person, or localized replacement is the central requirement.

Try the Krea2 prompt generator, open the Krea2 img2img guide, or use the Krea2 API reference.