Summary

Krea 2 offers a state-of-the-art open-source image generation model, emphasizing creative and stylistic exploration. It targets engineers, researchers, and power users seeking a highly aesthetic and performant solution, recognized as the #1 text-to-image model from an independent lab on Artificial Analysis. The project provides two distinct checkpoints optimized for a streamlined training and inference workflow.

How It Works

The core innovation is a dual-model architecture: Krea 2 RAW and Krea 2 Turbo. RAW serves as the base, undistilled checkpoint, ideal for fine-tuning and LoRA development due to its malleability. Turbo is an 8-step distilled checkpoint optimized for rapid, high-quality text-to-image generation. This design facilitates a powerful workflow where LoRAs trained on RAW can be seamlessly applied to Turbo for efficient, high-fidelity inference.

Quick Start & Requirements

Setup involves downloading safetensor files from Hugging Face and exporting them via OSS_RAW and OSS_TURBO environment variables. Inference is executed using the uv package manager with commands like uv sync and uv run inference.py. The model supports generating images up to 2048x2048 resolution. Key resources include API Docs, Hugging Face model pages, a Technical Blog, and Prompting/Safety Guides.

Highlighted Details

• Claimed as the "#1 text-to-image model from an independent lab on Artificial Analysis".
• Dual-model