Krea 2 AI Image Generator for Aesthetic Control

Krea 2 is Krea AI’s first image foundation model built completely from scratch. Krea’s technical report, published June 23, 2026, describes it as a 12.9-billion-parameter single-stream diffusion transformer that uses Qwen3-VL as its text encoder. It focuses on aesthetic diversity and creative control, combining text to image generation with prompt expansion and visual style references.

Krea 2 is designed to explore how an image looks and feels, not only whether it contains the requested objects. Its published strengths include expressive visual range, style references, moodboards, and adjustable creativity. Krea reports a top-ten placement on the Artificial Analysis text to image leaderboard, and released the model weights openly instead of keeping them behind an API.

A style reference is an image used to guide palette, texture, lighting, composition, or atmosphere. Krea’s user guide documents that several references can be combined, with each one influencing the result at a strength you set. Reach for them when the aesthetic you want is easier to show than to describe in words.

Creativity controls how literally Krea 2 follows a prompt. Raw and low settings stay closer to explicit instructions; medium balances fidelity and invention; high gives the model more freedom to add aesthetic decisions. Lower settings suit client work with fixed requirements, higher settings suit open-ended exploration where a surprising result is welcome.

Krea 2 Medium is positioned for faster, more cost-efficient creative exploration, especially expressive and illustrative work. Krea describes Large as more than twice the size of Medium, with stronger photorealism, raw texture, motion blur, grain, and cinematic looks. Medium is the practical default; Large earns its extra cost when photographic detail is the point.

Krea 2 Turbo is the speed-optimized, distilled Krea 2 variant. Krea’s technical report describes it as an eight-step checkpoint that renders 2K images in roughly two seconds on a modern GPU, with classifier-free guidance switched off. Krea published open weights for Turbo and the undistilled Raw checkpoint on Hugging Face in June 2026. Availability, inputs, pricing, and output terms still depend on the provider used for generation.

Start with the subject and purpose, then specify composition, camera or viewpoint, lighting, materials, color, mood, and medium. Krea 2 applies prompt expansion, so a short prompt gets elaborated for you — convenient for ideation, but it hands the model decisions you may want to make yourself. State what matters explicitly, and add references when the aesthetic is easier to show than describe.

Krea 2 supports moodboard-led direction through compatible generation interfaces. A moodboard groups references around a shared palette, texture, atmosphere, or composition so the model follows their collective intent rather than copying any single image. This is the practical way to hold one consistent look across a series of generations.

Commercial rights depend on the model provider and the plan used for generation. Krea released the Raw and Turbo weights on Hugging Face under its own community license rather than a standard open-source license, so self-hosted use carries terms worth reading in full. Review the applicable provider terms and this workspace’s terms before using an output in paid media, client work, or products.

No. krea-2.pro is an independent website and is not affiliated with, endorsed by, or sponsored by Krea AI. Krea and Krea 2 are referenced only to identify the model and related workflows. Model facts on this page are taken from Krea’s official technical report, product documentation, and published model cards, all linked below.