Abiray/Krea-2-Turbo-FP8-NVFP4

Krea 2 Turbo - FP8 & NVFP4 Safetensors

This repository contains optimized low-bit precision weights (.safetensors) of Krea 2 Turbo, the 12-billion parameter Diffusion Transformer (DiT) text-to-image model developed by Krea.ai.

These variants are tailored for optimized inference pipelines (such as ComfyUI) to drastically reduce VRAM utilization while preserving generation speeds and semantic visual accuracy.

---

📂 Available Files & Dimensions

File Name	Precision Type	Size	Description
krea2_turbo_fp8_scaled.safetensors	FP8 (Scaled)	~13.1 GB	8-bit floating-point format. Offers a near-perfect balance between high visual fidelity and reduced memory consumption on modern GPUs.
krea2_turbo_nvfp4.safetensors	NVFP4 (NVIDIA FP4)	~7.67 GB	Ultra-low 4-bit precision scaling tailored for compatible NVIDIA architectures. Ideal for maximizing generation performance on hardware with tighter VRAM limits.

---

⚙️ Usage Guide (ComfyUI)

These files are designed to drop directly into your modern diffusion workflow managers like ComfyUI.

1. Download your preferred model file (fp8_scaled or nvfp4).
2. Place the file inside your local directory structure:

   ComfyUI/models/diffusion_models/