Krea 2 (K2)

Krea 2 - image generation model from Krea.

---

API Docs • Hugging Face

This is the official repository for open version of Krea 2, an image model trained from scratch focused on creative and stylistic exploration. The repository contains minimal inference code and instructions to run the model.

Krea 2 ships as two models. Krea 2 RAW is the base model. It's a pretrained checkpoint with no distillation, so it's diverse and highly malleable, and it's what you should use for fine-tuning, post-training, and LoRA training. Krea 2 Turbo is an 8-step distilled checkpoint built for fast, high-quality text-to-image.

The two models are designed to work together. You train LoRAs on RAW and apply them on Turbo, and the LoRAs trained on RAW will work well on Turbo. We highly recommend using RAW for training LoRAs and applying them on Turbo for inference.

Setup

uv sync

Both Raw and Turbo safetensor files are available on Hugging Face. After downloading the checkpoints, set the OSS_RAW and OSS_TURBO environment variables to the paths of the downloaded files.

export OSS_RAW=...
export OSS_TURBO=...

Usage

The following commands run inference using the two available checkpoints with recommended settings.

Raw (oss_raw)

The base undistilled model. Use the full sampler with classifier-free guidance: The model has been trained to generate upto 1k resolution.

uv run inference.py "a fox walking in the snow" \
    --checkpoint oss_raw --steps 52 --cfg 3.5

Turbo (oss_turbo)

Distilled for few-step sampling — run with 8 steps and CFG disabled. The model can generate images from 1k ~ 2k resolution.

uv run inference.py "a fox walking in the snow" \
    --checkpoint oss_turbo --steps 8 --cfg 0.0 --mu 1.15 --width 2048 --height 2048

Options

Documentation

• Prompting Guide
• Safety Guide

FAQ

Which model I should use?

Use the Turbo model for fast inference with high quality results. The Raw model is an undistilled checkpoint without any step / cfg guidance distillation and posttraining. It is a highly finetunable base model that can be used to train LoRAs for the Turbo model as well as posttraining research. In short, TRAIN on Raw and RUN on Turbo.

What license is this model released under?

Both model weights are under our community license with permissive use. To purchase a commercial license, please contact us at opensource@krea.ai.