Krea launches Krea 2, a new foundation model family featuring a 12.9B-parameter DiT backbone and high-performance Elo scores across Large, Medium, and Turbo variants.
On May 1, 2026, Krea officially announced the release of Krea 2, a sophisticated family of image generation models designed to push the boundaries of aesthetic fidelity and creative control. Moving beyond their previous collaborative work, such as the FLUX.1 Krea [dev] co-developed with Black Forest Labs, Krea has now transitioned into a provider of proprietary foundation models built from the ground up.
For developers and AI engineers, Krea 2 represents a significant shift from simple prompt-to-image tools toward a robust, scalable creative suite. The release introduces a tiered model architecture—Large, Medium, and Medium Turbo—allowing users to balance computational cost against visual complexity and prompt adherence.
The technical backbone of Krea 2 is a 12.9B-parameter Diffusion Transformer (DiT) architecture. This design choice allows for superior scaling laws compared to traditional U-Net architectures, enabling the model to better capture complex spatial relationships and intricate textures.
A standout feature for professional workflows is the new batch variation control. This allows engineers to programmatically dictate the level of visual cohesion within a generated set. You can instruct the model to maintain strict stylistic consistency across a batch or allow for high-entropy exploration of a single prompt concept.
The model family also includes a 'Raw' variant specifically optimized for users who require minimal post-processing intervention, providing a clean slate for downstream fine-tuning or advanced image-to-image pipelines.
Krea 2 has demonstrated industry-leading performance in qualitative human preference testing. According to Artificial Analysis Elo scores, the Krea 2 Large model leads the family with a score of 1035, significantly outperforming the Medium variant (1014). The Medium Turbo model, released in June 2026, sits at 1021, offering a high-performance middle ground for latency-sensitive applications.
These scores reflect the model's ability to handle complex prompt instructions and maintain high aesthetic standards without the 'uncanny valley' effects often seen in smaller models. The evaluation methodology relies on blind human preference testing, where users rank outputs from Krea 2 against competing state-of-the-art models based on composition, lighting, and prompt adherence.
Krea has implemented a tiered pricing structure based on the specific model variant used. This allows developers to optimize their operational expenditure (OpEx) by selecting the appropriate model for their specific use case, whether it be high-fidelity production or rapid prototyping.
Pricing is calculated per 1,000 images generated. The Large model, intended for premium, high-detail production, is priced at $60 per 1k images. The Medium model offers a balanced approach at $30 per 1k images, while the Medium Turbo variant is optimized for speed and cost-efficiency at $15 per 1k images.
The Krea 2 family is uniquely suited for several high-end creative and technical workflows. In the realm of game development and concept art, the batch variation control allows artists to generate hundreds of consistent character or environment assets in a single pass.
For marketing agencies and advertising tech, the Medium Turbo variant provides the low-latency generation required for real-time creative tools and interactive web experiences. Additionally, the high Elo scores make it an ideal choice for high-fidelity asset generation in RAG-based visual workflows, where the quality of the retrieved or generated visual must match the complexity of the textual query.
Developers can access Krea 2 through the official Krea AI platform and via their dedicated API endpoints. The platform supports integration with popular workflows, including ComfyUI, making it easy to plug Krea 2 into existing node-based generative pipelines.
To begin, engineers can sign up at krea.ai to access the API documentation, SDKs, and playground environments. The API is designed for high throughput, supporting large-scale batch processing for enterprise-level applications.
API Pricing — Output: $60 per 1k images (Large), $30 (Medium), $15 (Medium Turbo) / Context: Pricing is based on image count per tier