Krea 2 is the first in-house foundation image model by Krea, the US company behind the Krea creative platform. On 22 June 2026 Krea released the weights in two variants: Krea 2 Raw, an undistilled base for fine-tuning and LoRA training, and Krea 2 Turbo, a distilled version for fast generation. The core idea: train on Raw, generate with Turbo. This page collects the key facts, notes from running it locally, and a critical assessment.
Key facts
| Developer | Krea (Krea.ai, Inc., USA) |
| Open-weights release | 22 June 2026 |
| Size | approx. 12.9 billion parameters |
| Architecture | Single-stream Diffusion Transformer, trained from scratch |
| Text encoder | Qwen3-VL-4B (12 stacked hidden layers) |
| VAE | Qwen-Image VAE (Wan 2.1 VAE in stable-diffusion.cpp) |
| Steps (Turbo) | 8, CFG 1 |
| Licence | Krea 2 Community License |
| Variants | Raw, Turbo |
What sets it apart
- Aesthetic range: Krea built the model explicitly for stylistic diversity. Where many models converge on a polished default look, Krea 2 produces visibly different results across attempts.
- No AI images in pretraining: according to the technical report, the team deliberately excluded AI-generated images, because even small amounts push models towards the generic “AI look”.
- Unconventional styles preserved: quality filters only remove extremely poor images. Even material such as low-resolution CRT footage was kept as a legitimate style.
- Speed: Turbo generates 2K images in around two seconds on fast hardware.
The price is size: at 12.9 billion parameters, the transformer alone takes about 26 GB in bf16 – roughly twice Z-Image.
Running it locally with stable-diffusion.cpp
stable-diffusion.cpp supports Krea 2 natively, including on AMD graphics cards via the Vulkan backend. Three files are needed:
- Diffusion model: e.g.
krea2_turbo-Q4_K_M.gguf(Q4 quants approx. 7–8 GB) - Text encoder:
Qwen3-VL-4B-Instructin GGUF format – the VL version, not plain Qwen3 - VAE:
Wan2_1_VAE(in ComfyUI:qwen_image_vae)
Recommended settings: 8 steps, CFG 1.0. Even a card with only 4 GB of VRAM (e.g. a Radeon RX 6400) can run it if the weights stay in system memory and are streamed to the GPU in portions – slower, but it works:
sd-server.exe ^
--diffusion-model krea2_turbo-Q4_K_M.gguf ^
--llm Qwen3-VL-4B-Instruct-Q4_K_M.gguf ^
--vae Wan2_1_VAE_bf16.safetensors ^
--backend "diffusion=vulkan0,te=cpu,vae=cpu" ^
--params-backend "diffusion=cpu,te=disk,vae=cpu" ^
--flash-attn off,diffusion=on ^
--conv-direct auto,vae=on ^
--vae-tiling ^
--max-vram vulkan0=2.0 ^
--width 768 --height 768 ^
--steps 8 --cfg-scale 1LoRAs are best trained on Krea 2 Raw and then applied to Turbo. LoRAs for Flux or Z-Image are not compatible.
Prompting
Krea 2 was trained mostly on long, densely detailed captions, but deliberately also on short and medium-length prompts. It therefore handles both, but rewards precise descriptions:
- Subject: who or what, with concrete details.
- Setting: place, background, secondary figures.
- Light and camera: flash, daylight, lens, grain.
- Style and medium: photography, painting, illustration, era – this is where Krea 2’s range shows most.
Because seeds actually produce different results, it is worth generating several variants of the same prompt.
Open weights – with conditions
The inference code is released under Apache 2.0, but the weights are under the Krea 2 Community License. It permits use, modification and redistribution, and outputs belong to the user. Commercial use, however, is only free for individuals and companies with less than USD 1 million in annual revenue; above that, an enterprise licence is required.
There are further strings attached: anyone deploying the model must implement content filters, AI-generated outputs must be labelled where law or platforms require it, derived models must carry “Krea” in their name, and Krea can revoke the licence with 30 days‘ notice. Compared with Z-Image’s Apache 2.0, this is considerably more restrictive – open, but not free.
The question of training data
Krea’s technical report is unusually detailed about curation: billions of images, deduplication via perceptual hashes and embeddings, classifiers, and Wikipedia’s PageRank to make sure important concepts of world knowledge are represented. Captions combine vision-language models with OCR.
Where the images come from is answered only in general terms. The model card names publicly available data, data licensed from third-party providers and synthetic data generated through proprietary methods – which sits in some tension with the report’s statement that no AI images were used in pretraining; the synthetic portion presumably relates to later training stages. Which archives, photographers or platforms are included is not disclosed.
In the community, Krea 2 is regarded as one of the least censored open image models in years. That, too, points to a broad, lightly filtered dataset. As with almost every image model, the open ethical question remains on whose images this aesthetic diversity was learned.