Abiray/Krea-2-Turbo-GGUF

Krea 2 Turbo - GGUF Quantizations

This repository contains GGUF format quantizations of Krea 2 Turbo, a state-of-the-art 12-billion parameter Diffusion Transformer (DiT) text-to-image model developed by Krea.ai.

GGUF weights allow for efficient, low-memory inference across consumer hardware, CPUs, and Apple Silicon.

---

📂 Available Files & Sizes

File Name	Quantization Type	Size	Description
Krea-2-Turbo-Q3_K_M.gguf	Q3_K_M	~6.01 GB	Highly compressed, lowest memory usage.
Krea-2-Turbo-Q3_K_S.gguf	Q3_K_S	~6.01 GB	Small variation of 3-bit quantization.
Krea-2-Turbo-Q4_K_M.gguf	Q4_K_M	~7.49 GB	Balanced medium quantization, recommended baseline.
Krea-2-Turbo-Q4_K_S.gguf	Q4_K_S	~7.49 GB	Small variation of 4-bit quantization.
Krea-2-Turbo-Q5_K_M.gguf	Q5_K_M	~8.87 GB	High quality, great balance of speed and retention.
Krea-2-Turbo-Q5_K_S.gguf	Q5_K_S	~8.87 GB	Small variation of 5-bit quantization.
Krea-2-Turbo-Q6_K.gguf	Q6_K	~10.6 GB	Near-lossless representation of original weights.
Krea-2-Turbo-Q8_0.gguf	Q8_0	~13.7 GB	Extremely high precision, closest to full 16-bit.

---

🚀 Inference Guide

Using llama.cpp / stable-diffusion.cpp

You can run these GGUF text-to-image files directly using compatible implementations such as stable-diffusion.cpp.

# Example using stable-diffusion.cpp for a 1024x1024 image generation
./sd -m ./models/Krea-2-Turbo-Q4_K_M.gguf \
  -p "a red fox sitting in fresh snow, golden hour, photorealistic" \
  --steps 8 --cfg 0.0 -w 1024 -h 1024 -o output.png