Krea 2 Turbo is the open-weights, distilled checkpoint of Krea 2 - Krea AI's first foundation image model, trained from scratch with a focus on how an image feels rather than just what it contains. Turbo is an 8-step build distilled from Krea 2 RAW and optimized for fast, high-quality inference at high resolution.
Originally released by Krea AI on Hugging Face. All credit for the model goes to the Krea team and the contributors listed below. Civitai is hosting a mirror so creators can generate with it on-site - please head to the original repo for weights, the technical report, and to support the project directly.
Built by
Krea AI, with Sangwu Lee, Erwann Millon, Le Zhuo, Matthew Newton, Andrei Filatov, Abhinay Devarinti, Dazhi Zhong, Avram Djordjevic, Gabriel Menezes, Will Beddow, Titus Ebbecke, Mihai Petrescu, Owen Fahey, Gian Saß, Felix Gil, and Victor Perez.
8-step distilled for speed
Turbo bakes guidance into the weights, so it runs without classifier-free guidance - Krea's recommended settings are 8 steps at CFG 0.0 with a constant timestep shift (mu) of 1.15. It generates at 1k to 2k resolution and supports outputs up to 2048x2048, making it the build to reach for when you want Krea 2 aesthetics at a fraction of the compute.
RAW-compatible LoRAs
Turbo is compatible with LoRAs trained on Krea 2 RAW, so the standard workflow is to train on RAW and run inference here. You can stack community LoRAs on top during on-site generation.
On-site generation
You can generate with Krea 2 Turbo directly on Civitai and add LoRAs on top. Weights are available upstream on Hugging Face under Krea's community license with permissive use; commercial licensing is available through [email protected].
Links
- Weights: huggingface.co/krea/krea-2-turbo
- Technical report: krea.ai/blog/krea-2-technical-report
- Product: krea.ai
Description
FAQ
Comments (9)
Note: at least at the moment, you should NOT set the CFG to 0 on the CivitAI generator, even though the description says so and the generator defaults to it. You will get a completely random image. CFG 1 seems to work just fine.
23 credits for 8 steps is crazy
I used the lightest FP8 version on a system with an RTX 3060 (12GB) and 32GB of RAM. Generation took 44 seconds with settings of 8 steps and a CFG scale of 1. While the heavier 12GB models also run, I didn't notice any difference in generation speed or image quality in my setup. I've posted the resulting image as well.https://civitai.red/posts/29494120
That's good because I used a 4-bit GGUF quant of the model on my 6GB GPU with 16GB of RAM and it dosn't take as much vram as Flux 2 Klein even when generating at 2K resolution which is amazing
Update: I’ve settled on the int8 version. It seems to be stable on the RTX 3000 series.
Theres 3 versions of FP8... can we have a rundown of whats the difference between them by size?
It's just the site hasn't proper labels for quants. In fact that are fp8, mxfp8 and nvfp4 quants
Please remove the recommendation for CFG 0.0 from the model description. That's complete misinformation as far as ComfyUI is concerned. You're supposed to set it to 1.0 instead.
I understand it says 0.0 in the official model release, but you just get completely random images in ComfyUI (which is kind of cool because they're very coherent, but probably not what you want).
For any serious LORA trainers, it saddens me to announce that after training 500 images based on various breast sizes (captioned with an even number of: flat chest, small breasts, medium breasts, large breasts), Krea2 CANNOT produce a flat chest and is hit or miss with breast sizes.
Unfortunately, it's on the same level as Z-Image. And for anyone who cares, Anima is still the only modern checkpoint that will identify breast sizes across characters.