Krea Releases 12B Image Weights for 2-Second 2K Generation

The interesting part of Krea's open-weight release is not just the speed claim -- it is how the two-checkpoint structure reflects a deliberate product strategy. On June 22, 2026, Krea published weights for Krea 2 Raw and Krea 2 Turbo on Hugging Face. As VentureBeat reported, Turbo targets enterprise-grade image generation in roughly two seconds at native 2K resolution on consumer hardware, using eight inference steps with classifier-free guidance disabled.

The technical foundation is a 12-billion-parameter single-stream diffusion transformer using grouped-query attention, with Qwen 3 VL as the text encoder and the Qwen Image VAE. Training moved through multiple stages -- pretraining at progressively higher resolutions, supervised finetuning, preference optimization, and reinforcement learning using reward models for aesthetics, prompt-following, text rendering, and artifact detection. That multi-reward RL stage mirrors the post-training approach that has improved large language models, applied here to visual quality rather than instruction-following.

The two checkpoints serve distinct audiences. Krea 2 Raw is the undistilled base designed for fine-tuning and LoRA training -- the intended raw material for developers building specialized variants for domains like automotive, architecture, or advertising. Krea 2 Turbo is the distilled, speed-optimized endpoint. Krea says LoRAs trained on Raw are designed to transfer well to Turbo, which would let developers get both fine-tuned control and fast inference without retraining from scratch. Krea 2 placed second among independent labs on the Artificial Analysis text-to-image leaderboard, according to the company's own technical report -- a benchmark position worth verifying as third-party evaluations accumulate.

The license is the business model made explicit. Individuals and small teams can use the weights commercially for free, but organizations reportedly requiring more than 50 seats, SSO integrations, SLAs, or custom data agreements must move to a paid Enterprise tier. All users must implement technical safeguards against generating illegal materials and non-consensual intimate imagery. Krea is using open weights as a distribution strategy: let developers and small studios build on and validate the model, then convert at the organizational scale boundary.

What the current reporting does not give you is Enterprise pricing, a clear hardware floor for the two-second generation claim, or quantified evidence of how well LoRA transfer from Raw to Turbo actually performs at 2K resolution -- all gaps that matter before committing a production pipeline to the model.