krea-ai/krea-2

Official inference code for Krea 2

What it solves

Krea 2 provides a high-aesthetic, open-source text-to-image generation system designed for creative and stylistic exploration. It addresses the need for a model that is both high-quality for end-users and highly flexible for developers who want to fine-tune the model for specific styles.

How it works

The system consists of two complementary models:

  • Krea 2 RAW: A base, undistilled pretrained checkpoint. Because it lacks distillation, it is diverse and malleable, making it the ideal foundation for fine-tuning, post-training, and LoRA training.
  • Krea 2 Turbo: An 8-step distilled checkpoint optimized for fast, high-quality image generation.

Users are encouraged to train LoRAs on the RAW model and then apply those LoRAs to the Turbo model for efficient inference.

Who it’s for

This project is for digital artists, AI researchers, and developers who need high-resolution (up to 2k) image generation and the ability to customize the model's output via fine-tuning.

Highlights

  • Dual-Model Architecture: Separate models for training (RAW) and fast inference (Turbo).
  • High Resolution: Supports generation from 1k up to 2k resolution.
  • Performance: Ranked as a top text-to-image model from an independent lab on Artificial Analysis.
  • Flexible Ecosystem: Compatible with ComfyUI, SGLang, and various fine-tuning toolkits like Huggingface Diffusers and Kohya.

Related

  • Dispatch
    Krea 2 Release: SOTA Open-Weights 12B Image Model

    Krea has released Krea 2, a state-of-the-art 12B parameter open-weights image model available in Turbo and RAW versions for high-performance text-to-image generation.

  • Project
    lbouaraba/comfyui-krea2edit

    A ComfyUI node pack for instruction-based image editing using Krea 2, enabling high-fidelity identity preservation through dual latent and semantic conditioning.

  • Project
    ModelTC/Minimax-H3-Turbo

    A collection of distilled LoRA checkpoints and workflows for MiniMax-H3 that enable high-quality video and audio generation in as few as 4 to 8 inference steps.

  • Project
    huggingface/finetrainers

    A library for the accessible training and fine-tuning of diffusion models, focusing on memory-efficient text-to-video and text-to-image generation.