Flux.2 Klein 9B — KV Consistency Edit [Fast]

Flux.2 Klein 9B — KV Consistency Edit [Fast] is an optimized, image-to-image editing suite designed around the advanced black-forest-labs/FLUX.2-klein-9b-kv base model. By incorporating a dedicated pipeline patch (flux2_klein_kv.patch) alongside the dx8152/Flux2-Klein-9B-Consistency LoRA adapter, this suite yields identity-preserving, context-consistent adjustments over input images in a fraction of standard rendering windows.

Operating entirely on custom CUDA setups, the suite offers text-guided image manipulation—such as seasonal swaps, complex structural relighting, and high-fidelity texture enhancement—with zero reliance on external APIs.

Key Features

• KV Attention Modification: Integrates a core structural patch directly over local diffusers modules via subprocess initialization to enable Key-Value consistency mechanisms.
• Klein-Consistency LoRA Engine: Leverages the Flux2-Klein-9B-Consistency adapter at a unified weight scale ($1.0$), ensuring structural traits and composition stay fixed during inference.
• Intelligent Resolution Adaptation: Parses active Gradio Gallery inputs and downscales or upscales dimensions to match native training boundaries, snapping layouts to multiples of 8.
• Unified ZeroGPU Workflow: Features memory cleanup utilities (gc.collect() and torch.cuda.empty_cache()) coupled with the @spaces.GPU context executor to provide multi-step editing without encountering out-of-memory overheads.

Repository Structure

├── examples/
│   ├── 1.jpg
│   ├── 2.jpg
│   ├── 3.jpg
│   └── 4.jpg
├── app.py
├── flux2_klein_kv.patch
├── LICENSE.txt
├── pre-requirements.txt
├── pyproject.toml
├── README.md
├── requirements.txt
└── uv.lock

Installation and Requirements

To run Flux.2 Klein 9B — KV Consistency Edit locally, ensure you possess an appropriate Python configuration compiled with heavy weight execution backends. A modern CUDA-enabled GPU is required.

This repository specifically relies on PyTorch 2.11.0 and CUDA 13.0 (--extra-index-url https://download.pytorch.org/whl/cu130).

Running with uv (Recommended)

uv is an ultra-fast Python package and project manager written in Rust, ensuring rapid virtual environment synchronization and reproducible execution.

Step 1 — Install uv

• macOS / Linux: curl -LsSf https://astral.sh/uv/install.sh | sh
• Windows: powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

Step 2 — Clone the repository

git clone https://github.com/PRITHIVSAKTHIUR/flux-klein-kv-edit-consistency-fast.git
cd flux-klein-kv-edit-consistency-fast

Step 3 — Initialize the project and install dependencies This will automatically parse the uv.lock and requirements.txt to fetch the correct PyTorch 2.11.0 + cu130 wheels.

uv sync

Step 4 — Run the script

uv run app.py

Standard PIP Installation

1. Install Pre-requirements Ensure your local system package manager is upgraded:

pip install pip>=26.0.0

2. Install Core Dependencies Install the primary deep learning stack, diffusion utilities, and ecosystem structures. Place these in a requirements.txt file and execute pip install -r requirements.txt.

--extra-index-url https://download.pytorch.org/whl/cu130

git+https://github.com/huggingface/transformers.git@v4.57.6
huggingface-hub
gradio==6.16.0
torch==2.11.0
opencv-python
sentencepiece
torchvision
torchaudio
accelerate
omegaconf
termcolor
diffusers
kernels
imageio
hf_xet
spaces
pyyaml
pillow
numpy
peft
ftfy
av

Usage

After setting up your environment and ensuring your dependencies are installed, launch the application by executing the primary module script:

python app.py

The script will trigger a safety lookup for your environment, parse flux2_klein_kv.patch, and hot-patch your local site-packages diffusers library. It will then download and cache the 9B model weights along with the consistency LoRA layers. Once initialized, a local web server interface will be exposed (typically at http://127.0.0.1:7860/).

1. Input Asset: Upload one or more reference images to the Gradio input gallery. Leaving the panel completely clear switches the underlying generation path back into text-only mode.
2. Define Modification: Enter your specific adjustments in the text box (e.g., "Transform the scene into a snowy winter day").
3. Advanced Tweaks: Expand the Advanced Settings menu to scale inference steps, lock baseline seeds, or manually enforce structural dimension ratios.
4. Compile: Click Edit Image to launch the CUDA workspace worker thread and review the consistency-matched result.

License and Source

• License: Apache License 2.0
• GitHub Repository: https://github.com/PRITHIVSAKTHIUR/flux-klein-kv-edit-consistency-fast.git