⚡ FLUX.2 Klein Local GUI (RTX 5090 Optimized)

A lightweight, high-performance local interface for FLUX.2 Klein (9B) by Black Forest Labs. This project is specifically optimized for NVIDIA RTX 5090 GPUs, utilizing PyTorch Nightly (CUDA 13.0) and bfloat16 precision for maximum speed.

🚀 Features

• Fully Portable: Embedded Python — no system Python installation required.
• Optimized for Blackwell: Built on PyTorch Nightly with CUDA 13.0 support for RTX 5090.
• Fast Generation: Pre-configured for the distilled "Klein" model (only 4 steps required).
• Dual Modes: Supports Text-to-Image and Image-to-Image.
• Memory Efficient: Implements CPU offloading to manage the 9B parameter model effectively.
• One-Click Installer: Includes a robust batch script for easy setup on Windows.
• Real-time Monitoring: Live CPU, RAM, and VRAM monitoring in the GUI.

📋 Prerequisites

• OS: Windows 10/11
• GPU: NVIDIA RTX 3090 / 4090 / 5090 (24GB+ VRAM recommended)
• Git: Required to install dependencies.
• Hugging Face Account: FLUX.2 is a gated model. You need to:
  1. Create a free account at huggingface.co.
  2. Accept the model license at FLUX.2-klein-9B.
  3. Generate an access token at HF Tokens.

🛠️ Installation

1. Download this repository.
2. Double-click install.bat.
3. The script will:
   • Download and set up an embedded Python environment (python_env/).
   • Install all dependencies (PyTorch Nightly, Diffusers, Gradio, etc.).
   • Ask you to log in to Hugging Face (required once to download the gated model).

No system Python and no virtual environment required.

Usage

Double-click start_FLUX2-KLEIN-9B_gui.bat.

Wait for the model to load (first run downloads approx. 52GB to model_cache/).

The GUI will open automatically in your browser (usually http://127.0.0.1:7860).

Recommended Settings for Klein

The Klein model is distilled, meaning it behaves differently than the base model:

• Steps: 4 steps is the sweet spot.
• Guidance Scale: Leave at 1.0.
• Resolution: 1024x1024 works best.

Troubleshooting

• 401/403 Client Error: You are not logged in or haven't accepted the model license.
  • Go to FLUX.2-klein-9B and accept the license.
  • Re-run install.bat and log in again, or run .\python_env\python.exe -m huggingface_hub.commands.huggingface_cli login.
• OOM (Out of Memory): Ensure you don't have other heavy GPU apps running. The app uses CPU offloading to save VRAM.
• Python environment not found: Make sure you ran install.bat before starting the app.

License

This project uses the FLUX.2 model. Please review the Black Forest Labs license for commercial usage restrictions.

🤝 Support

This is a free open-source project. I don't ask for donations. However, if you want to say "Thanks", check out my profile on Spotify. A follow or a listen is the best way to support me! 🎧