Krea 2 Medium is a foundation image generation model developed by Krea, officially introduced in May 2026. Built from scratch as a latent diffusion model, it is designed to prioritize aesthetic diversity and creative control over the generic "polished" look common in many contemporary generators. The model is trained to understand the visual language and atmospheric qualities of an image, such as film grain, material textures, and lighting styles, allowing for a wider range of expressive outputs.
The model's architecture supports several advanced features for granular control, including a Style Transfer system and Moodboard Conditioning. Users can provide one or more reference images, and the model extracts their aesthetic profile to apply to new generations. A tunable Creativity parameter allows for adjustment between literal prompt adherence and aesthetic interpretation, with higher settings adding more visual richness and depth.
Within the Krea 2 series, the Medium variant is the more efficient and faster option. It features heavier post-training compared to the Large variant, which provides greater stability and consistency across consecutive generations. This stability makes it particularly well-suited for illustrations, anime, painterly styles, and commercial graphic design workflows. It excels at rendering complex textures like iridescent surfaces, metallic reflections, and soft gradients.
Prompting Krea 2 Medium is most effective when using natural language descriptions combined with visual references. The model often makes intentional compositional and lighting choices—such as "golden hour" tones or specific depth-of-field—that serve the overall aesthetic even when not explicitly requested in the text prompt. It supports high-resolution outputs and is capable of generating detailed imagery of subjects ranging from humans and animals to architectural concepts.