Krea AI · 2026-06-23 · major
Krea 2 — open-weight 12B image model with 2-second Turbo variant
Krea 2 is a 12B Diffusion Transformer text-to-image model released as open weights in two variants: Raw for fine-tuning and Turbo, which generates 2K images in about 2 seconds.
Krea AI open-sources a 12B Diffusion Transformer image model with a Turbo variant that draws 2K in two seconds.
Key specs
| Parameters | 12B |
|---|---|
| Turbo speed (2 k) | ~2s |
Quick facts
| Maker | Krea AI |
|---|---|
| Parameters | 12B Diffusion Transformer |
| Variants | Krea 2 Raw (base) and Krea 2 Turbo (8-step distilled) |
| Text encoder | Qwen3-VL with multi-layer feature aggregation |
| License | Krea 2 Community License (Enterprise required above 50 seats) |
| Where to run | Hugging Face, GitHub, ComfyUI, FAL, AWS, GCP |
| Released | June 23, 2026 |
What is it?
Krea 2 introduces a 12B Diffusion Transformer trained from scratch for aesthetic-first text-to-image generation. Krea AI is publishing two checkpoints: Krea 2 Raw, the undistilled base for LoRA fine-tuning and post-training research, and Krea 2 Turbo, an 8-step distilled inference engine. Both ship as open weights under the Krea 2 Community License, with an Enterprise tier for firms above 50 seats.
How does it work?
Under the hood Krea 2 pairs a 12B dense DiT backbone with a Qwen3-VL text encoder using multi-layer feature aggregation, plus a Qwen Image VAE. The pipeline uses grouped-query attention with sigmoid-gated attention, SwiGLU MLPs, and lightweight timestep modulation. Training runs through pretraining at 256, 512, and 1024 pixels, midtraining, supervised finetuning, preference optimization, reinforcement learning, and timestep distillation for the Turbo checkpoint.
Why does it matter?
An independent lab shipping a 12B open-weight image model that draws 2K in 2 seconds tightens the gap with proprietary text-to-image APIs and the closed Black Forest Labs models. Because Raw is published, downstream studios and researchers can fine-tune Krea 2 for their own style without retraining from scratch — Krea cites Superside already routing about 80% of its production through Krea, and ComfyUI shipped same-day nodes for both checkpoints.
Who is it for?
Image generation researchers, downstream studios fine-tuning a base model, and ComfyUI workflow builders.
Frequently asked questions
- What is the difference between Krea 2 Raw and Krea 2 Turbo?
- Krea 2 Raw is the undistilled base checkpoint Krea trained from scratch; it is meant for LoRA fine-tuning and post-training research, not direct inference. Krea 2 Turbo is the same 12B model distilled to 8 steps for fast inference at 2K resolution, running in roughly 2 seconds on consumer hardware.
- Can I use Krea 2 commercially?
- Krea 2 ships under the Krea 2 Community License. Individuals and teams below 50 seats can use the weights for free, including for commercial work, but firms above 50 seats must take an Enterprise license. All users must implement safeguards against generating illegal material, non-consensual intimate imagery, CSAM, or defamatory content.
- How does Krea 2 compare to other open image models?
- Krea AI states Krea 2 ranks among the top 10 text-to-image models on the Artificial Analysis leaderboard and is the second-highest model from an independent lab, after Black Forest Labs. Krea built the 12B Diffusion Transformer from scratch rather than fine-tuning an existing checkpoint, and pairs it with a Qwen3-VL text encoder for stronger prompt following.
- Where can I run Krea 2 locally?
- Krea 2 weights are on Hugging Face at krea/Krea-2-Raw and krea/Krea-2-Turbo, with official inference code at github.com/krea-ai/krea-2. ComfyUI shipped day-one support, and the model also runs on FAL, AWS, and Google Cloud. The Turbo variant is the recommended starting point for inference; Raw is for fine-tuning.
Try it
huggingface.co/krea/Krea-2-Turbo (or pull github.com/krea-ai/krea-2)