Krea 2 Open-Weights 12B Text-to-Image Model Release

Krea 2 Open-Weights 12B Text-to-Image Model Release

Krea 2 releases open‑weight 12B text‑to‑image models

Krea has published the weights for two 12‑billion‑parameter text‑to‑image models—Krea 2 Turbo and Krea 2 RAW—along with a comprehensive technical report that explains data curation, architecture, training pipelines, and infrastructure. The release provides a rare open‑source glimpse into a production‑grade diffusion model and a permissive license for individuals and small businesses.

Two checkpoints target different use cases

  • Krea 2 Turbo is a guidance‑ and timestep‑distilled checkpoint designed for fast inference. It is already available in GGUF format for efficient CPU execution.
  • Krea 2 RAW is the un‑distilled checkpoint intended for research, fine‑tuning, and further experimentation. It is released at both a mid‑training and a post‑training stage, a practice uncommon in the image‑generation community.

"We are releasing two checkpoints at both the mid‑training and post‑training stage. This is rare in the image & multimedia community, so we can't help but feel proud of this release." – Diego Rodriguez, Co‑founder & CTO of Krea

Technical report covers end‑to‑end pipeline

The accompanying report (https://www.krea.ai/blog/krea-2-technical-report) details:

  • Data curation and captioning – how billions of image‑text pairs were collected, filtered, and annotated.
  • Model architecture – a diffusion backbone built on the Qwen‑3‑VL vision‑language encoder and a VAE decoder.
  • Post‑training and RL pipelines – reinforcement‑learning‑based fine‑tuning for style adherence and prompt expansion.
  • Infrastructure – the distributed training stack, storage layout, and cost‑optimisation strategies that enabled training a 12B model at scale.

Quality comparable to leading closed‑source models

Artificial Analysis benchmark results place Krea 2 on par with Nano Banana in text‑to‑image quality, despite being fully open‑source.

"We are on par with Nano Banana in terms of image quality as per Artificial Analysis text‑to‑image benchmarks." – Diego Rodriguez

Licensing and accessibility

Krea 2 is released under a permissive license that permits commercial use by individuals and small businesses. The models are hosted on Hugging Face (https://www.krea.ai/krea-2/huggingface) and the codebase is available on GitHub (https://www.krea.ai/krea-2/github).

Community reaction and future directions

  • Broad style capability – Commenters praised the "keep the manifold wide" philosophy, which aims for a model capable of many artistic styles rather than a limited set of presets.
  • Image‑to‑image frontier – Some users noted that while Krea 2 excels at text‑to‑image generation, the next major challenge is robust image‑to‑image editing, a capability where newer models like Nano Banana 2 and Images 2.0 are already advancing.
  • Technical depth appreciated – The in‑depth write‑up on training infrastructure was highlighted as a valuable resource for the community.

"Good to have more open weight models, and I really appreciate the in‑depth write‑up. I also like the ‘keep the manifold wide’ approach… the next frontier for image models lies in robust I2I."

How to get started

  1. Download the model – Choose Turbo for fast inference or RAW for research from the Hugging Face repository.
  2. Read the report – The technical report provides step‑by‑step details on data pipelines, model architecture, and training hyper‑parameters.
  3. Experiment – Use the provided inference scripts or integrate the model into existing diffusion pipelines. The GGUF‑converted Turbo checkpoint enables low‑resource CPU usage.
  4. Contribute – The open‑source license encourages community contributions, fine‑tuning, and downstream applications.

Bottom line: Krea 2 delivers state‑of‑the‑art open‑weight text‑to‑image generation with two distinct checkpoints, a thorough technical report, and a permissive license, marking a significant step toward democratizing high‑quality diffusion models.

Sources