Krea 2: The Open-Source Foundation Image Model for Creative Control

What Is Krea 2?

Krea 2 is Krea's first foundation image model, trained completely from scratch and released on May 12, 2026. Unlike many image models that build on top of existing architectures, Krea 2 was designed from the ground up with two primary goals: aesthetic diversity and creative control.

The Krea team identified a fundamental gap in AI image generation: most models are great at understanding what you want in an image, but they struggle with the "how it looks" part. They tend to default to a polished, safe, "AI look" that flattens artistic expression. Krea 2 was built to solve this problem.

The model is now open-source with weights released on Hugging Face, code on GitHub, and a permissive license for commercial and personal use. It's available through the Krea platform, API partners (Fal, Comfy, Runware), and directly via Nous Research's Hermes.

The Core Innovation: Style Transfer

What sets Krea 2 apart from every other image generation model is its advanced style transfer system. Rather than treating style as a vague word in a text prompt, Krea 2 lets you:

• Pass reference images to extract style components and apply them to your outputs with precision
• Combine multiple style references — blend styles together in adjustable proportions
• Control style intensity — from a light directional influence to a dominant visual force
• Adjust batch variation — keep images cohesive across a batch or push toward wider stylistic spreads

This means style is no longer a guessing game. You can pass a film photograph reference and get grainy, authentic results, or a clean studio shot reference and get polished output — the model responds to visual direction the way a human artist would respond to a mood board.

Model Variants: Medium vs Large

Krea 2 ships in two API variants, each targeting different use cases:

Krea 2 Medium has extensive post-training applied, making its outputs especially stable and consistent — ideal for commercial work where predictability matters. Krea 2 Large is more than twice the size with softer post-training, giving outputs a rawer, more textured, and flexible character that creative professionals prefer for experimentation.

Technical Architecture

The Krea 2 Technical Report (published June 23, 2026 by Sangwu Lee, Erwann Millon, Le Zhuo, and the full Krea research team) outlines their approach:

• Multi-stage training pipeline — spanning pretraining through post-training with careful curation at each stage
• Large-scale data infrastructure — built from scratch to curate a comprehensive pretraining dataset with broad world knowledge and style coverage
• Distributed training framework — enabling the model to learn from diverse, high-quality datasets
• Native 4K resolution — generation support for high-resolution output with over 1,000 style presets

The model weights are released on Hugging Face and the inference code on GitHub under a permissive license.

API & Developer Access

Krea 2 is available through multiple channels:

• Krea Platform — direct access via krea.ai/image/k2 with the full creative tool suite
• API Partners — Fal, Comfy, Runware for programmatic access
• Nous Research — integrated with Hermes since day 0
• Open Weights — self-host via Hugging Face weights and GitHub code

Why Krea 2 Matters

Krea 2 represents a shift in how image generation models approach creative exploration. Instead of producing polished but homogenized results, it's designed to be an exploratory medium — capable of spanning many aesthetics while giving creators the tools to navigate them deliberately.

The model is already being adopted by creative studios, marketing agencies, fashion brands, gaming companies, and architecture firms. Krea's research lab actively incorporates user feedback to evolve the model.

Related Articles

Top Open Source Image Generation Models — Complete ranking of image generation models for June 2026 Best AI Models for Hermes Agent — Local and cloud deployment options