Krea AI has open-sourced the weights of its flagship Krea 2 image generation model, releasing two checkpoints on Hugging Face for the developer and research community. The move puts enterprise-grade generative image capabilities into the hands of fine-tuners, researchers, and small-scale commercial developers at no cost — a significant milestone for the open image generation ecosystem.
Key Highlights
- Krea 2 is a 12 billion parameter Diffusion Transformer (DiT) — one of the largest open image models released to date
- Two variants: Krea 2 RAW (undistilled base for fine-tuning and LoRA training) and Krea 2 Turbo (8-step distilled for 2-second inference)
- Available on Hugging Face under the Krea 2 Community License
- Free for commercial use by companies with annual revenue under $1 million USD
- Enterprise license required for organizations with 50 or more seats
- Both checkpoints work with the Hugging Face Diffusers library and ComfyUI
Technical Architecture
Krea 2 is built on three components: a Qwen Image VAE, a 12B dense DiT backbone, and a Qwen3-VL text encoder with multi-layer feature aggregation. The Qwen3-VL text encoder gives the model strong natural language understanding, allowing it to handle complex, detailed prompts across professional domains.
The RAW checkpoint runs at 52 steps with a guidance scale of 3.5 at 1024x1024 resolution. It is the undistilled base model — not optimized for direct inference, but designed for researchers and developers building custom models through fine-tuning or LoRA training. The Turbo checkpoint distills the base into 8 steps, delivering generation in approximately 2 seconds without significant quality loss. LoRAs trained on RAW transfer effectively to Turbo, giving developers a direct path from training to fast production deployment.
Licensing Structure
Rather than a standard permissive license, Krea 2 ships under a custom community license:
- Free for small companies and individuals: commercial use is permitted for any organization with annual revenue under $1 million USD
- Enterprise license required for larger teams: organizations with 50 or more seats must contact Krea at opensource@krea.ai
- Content safeguards are mandatory for all users: implementations must include technical filters against non-consensual intimate imagery (NCII), CSAM, and other prohibited content categories — regardless of company size
Getting Started
Krea 2 integrates with the Hugging Face Diffusers library. Install the latest version from source, then run inference with Turbo for fast results:
pip install git+https://github.com/huggingface/diffusers.gitimport torch
from diffusers import Krea2Pipeline
pipe = Krea2Pipeline.from_pretrained(
"krea/Krea-2-Turbo",
torch_dtype=torch.bfloat16
).to("cuda")
image = pipe("A modern building facade with glass and steel under golden hour light").images[0]
image.save("output.png")For fine-tuning workflows, load krea/Krea-2-Raw instead and use standard LoRA training recipes compatible with DiT architectures. The full inference code and training guides are available on GitHub at krea-ai/krea-2.
Performance
In a four-model style-transfer benchmark conducted by Contra Labs, Krea 2 scored within 0.14 points of GPT Image 2 on Style Fidelity — a strong result for an open-weight model competing with a proprietary commercial system. The model targets professional creative workflows across several verticals:
- Architecture and real estate: materials, lighting, interior renderings, and facade visualization
- Advertising and editorial: campaign directions, product concept frames, and brand imagery
- Design systems: packaging, interface mockups, branding moodboards
- Automotive: configurator concepts and editorial launch visuals
Background
Krea AI originally launched Krea 2 as a proprietary API model on May 12, 2026. The company's platform has grown to over 30 million users across 191 countries. The open weights release on June 22, 2026 extends access beyond the API and positions Krea 2 as a foundation model for the broader image generation developer ecosystem.
What's Next
The RAW-to-Turbo training path lowers the barrier for building production-ready specialized image services. Expect a LoRA ecosystem to develop rapidly — similar to what emerged around Stable Diffusion and Flux — with fine-tunes targeting specific industries, artistic styles, and regional aesthetics. Both checkpoints are already available in ComfyUI for node-based workflow integration.
For developers building image products in the MENA region — where Arabic script rendering, Islamic geometric aesthetics, and regional visual sensibilities are key — Krea 2 RAW provides a capable 12B foundation for fine-tuning on locally relevant datasets.
Source: Krea AI