FLUX tiled upscaler

A minimal tiled-image upscaler that uses FLUX1.0-dev and ControlNet-Union. This upscaler uses both tile and low quality controlnets to upscale images.

Original	Upscaled

Installation

Tested with python 3.10, torch 2.1 and CUDA 11.8.

cd flux-tiled-upscaler
pip install -r requirements.txt

export HUGGINFACE_TOKEN="YOUR_TOKEN"
python scripts/download_models.py

You can then run the upscaler as follows:

# Upscale a single image by 2x
python run_upscaler.py --input_path examples/camp.jpg --output_folder results --upscale_factor 2

# Upscale a folder of images by 4x
python run_upscaler.py --input_path examples/ --output_folder results --upscale_factor 4

# Adjust quality settings
python run_upscaler.py --input_path examples/portrait.jpeg --num_inference_steps 12 --tile_control_scale 0.45 --low_quality_control_scale 0.7

Alternatively, you can launch the Gradio demo to use it interactively as follows:

# Run the interactive Gradio web demo
python gradio_demo.py

License

This codebase is licensed under the MIT license. Please refer to the model pages of FLUX and ControlNet-Union for their licenses.

From neuralwork with ❤️