Krea 2 Releases Open Weights for Raw and Turbo Models

What's new

Krea AI released the technical report and open weights for Krea 2 (K2), a text-to-image foundation model designed for high-fidelity creative work. The release includes two primary versions: K2 Raw and K2 Turbo. K2 Raw is optimized for photographic realism and high-resolution output, while K2 Turbo utilizes a 4-step distilled architecture to enable real-time image generation. Both models are built on a 1.2 billion parameter DiT (Diffusion Transformer) architecture and support a native resolution of 1024x1024 pixels.

The Krea 2 update introduces a specialized training methodology that prioritizes aesthetic quality over simple prompt adherence. According to the technical report, the models were trained on a curated dataset of 20 million high-quality images, focusing on lighting, composition, and texture. K2 Raw specifically targets the elimination of the plastic-like skin textures often found in AI-generated imagery, providing a more cinematic look for concept art and storyboarding.

How it fits your workflow

Krea 2 serves as a versatile tool for pre-production and asset creation, offering a middle ground between the massive parameter count of Flux.1 and the speed of SDXL Turbo. For concept artists, K2 Raw provides a viable alternative to Midjourney v6 when local control and privacy are required. The open-weights nature of Krea 2 allows studios to host the model on their own hardware, ensuring that proprietary character designs or set layouts remain secure within their internal pipeline.

The K2 Turbo model is particularly useful for live art sessions and rapid prototyping. Similar to the real-time capabilities of Leonardo.ai or the LCM (Latency Consistency Models) used in Stable Diffusion, K2 Turbo allows editors to see visual changes instantly as they adjust their text prompts. This immediate feedback loop is essential for directors who need to iterate on visual styles during a briefing or for VFX artists who need to generate quick matte painting foundations. By providing the weights openly, Krea enables the community to develop custom LoRAs and ControlNet adapters, which will likely expand its utility for specific tasks like architectural visualization or character consistency.

What it costs / how to try it

Krea 2 is available for use on the Krea AI web platform, and the model weights for K2 Raw and K2 Turbo have been released on Hugging Face under a research-friendly license. Users can download the weights to run the models locally on compatible GPU hardware or via cloud-based inference services.