TL;DR
Krea releases Krea 2 Raw and Turbo as open-weight diffusion transformers with 2-second generation times and customizable licensing for enterprises.
Key Points
- Krea 2 Turbo generates 1024x1024 images in ~2 seconds on consumer hardware using 8 inference steps
- 12-billion parameter single-stream Diffusion Transformer architecture with Grouped-Query Attention and SwiGLU MLPs
- Dual-checkpoint strategy: Raw (undistilled, for fine-tuning LoRAs) and Turbo (distilled, for production inference)
- Community License permits free commercial use up to 50 seats; enterprises require custom paid licensing
Why It Matters
Open-weight image models at this speed enable developers to build real-time creative workflows without proprietary API dependencies. The Raw/Turbo bifurcation lets teams fine-tune on unconstrained latent spaces then deploy at scale—critical for production pipelines where both training flexibility and inference speed matter.
Source: venturebeat.com