Krea opens up weights for its Krea 2 image models

Krea 2 is an aesthetic-focused image generation family that you can now run and train on your own hardware.

Instead of dropping a single compromise checkpoint, Krea released two distinct open-weight models on Hugging Face: Krea 2 Raw and Krea 2 Turbo.

Train on Raw, run on Turbo

Fine-tuning heavily distilled models is usually painful, while running raw base models locally eats up too much compute. Krea solves this by splitting the pipeline cleanly in two.

Krea 2 Raw is an undistilled checkpoint captured mid-training. It preserves wide aesthetic diversity and high malleability, making it the designated base for full fine-tuning, post-training, and training LoRAs inside setups like ComfyUI.

Krea 2 Turbo is the companion inference checkpoint. It is distilled down to an 8-step generation process designed for fast, polished text-to-image output.

Here is the clever part: you train your LoRAs on Raw, and then load those exact weights directly onto Turbo for daily generation. According to Krea, LoRAs trained on the Raw base express strongly on the 8-step Turbo model, giving you easy training without sacrificing fast generation speeds.

Under the hood

Underneath, Krea 2 is built on a 12B dense DiT backbone.

It pairs that diffusion core with a Qwen3-VL text encoder using multi-layer feature aggregation to interpret prompt nuances, alongside a Qwen Image VAE for visual encoding.

Krea highlights strengths in rendering materials, lighting, interior layouts, architectural facades, and early-stage spatial concepts.

Both checkpoints are available on Hugging Face. Deployment paths are split into community and enterprise licenses, with dedicated commercial terms for production rollouts.

Why it matters

If you build visual workflows or train custom styles in ComfyUI, Krea 2 is definitely worth a spin. A 12B dense DiT is beefy, but the Raw-to-Turbo handoff solves a real headache: you get a flexible training foundation without being forced to run slow 50-step inference on every single prompt.

Releasing an undistilled mid-training checkpoint is a welcome move that treats open-source developers like real collaborators rather than just passive consumers. For teams building specialized design pipelines, this might quickly become a favorite base to customize.