LoRA Image Preprocessor

A Python tool to preprocess images for LoRA training, with optional face cropping, background removal, and captioning.

Features

• Face-aware square cropping
• Background removal with rembg
• LLaVA captioning for full images
• Blendshape detection for facial expressions
• Easy CLI interface

Installation

git clone https://github.com/baba-yaga/lora-image-preprocessor.git
cd lora-image-preprocessor
pip install -r requirements.txt

Running

The main script is lora_image_preprocessor/main.py. You can run it from the project root directory.

CLI Options

• --input_dir: (Required) The directory containing the images to process.
• --output_dir: The directory base name where the processed images and captions will be saved. The actual name will be appended with the resoltion of the outputs and other flags: face and/or no_bg, if given. Defaults to the input directory.
• --face_only: If specified, the script will crop the image to the detected face. Note that in images where both eyes are not visible, or in cartoon-like images, the face is usually not detected.
• --remove_bg: If specified, the background of the images will be removed.
• --blur_bg: If specified, the background of the images will be blurred using a Gaussian kernel.
• --resolution: The target output resolution for the images (default: 512).
• --output_format: The output image format (e.g., png, jpg) (default: png).
• --face_crop_padding: The padding factor for face cropping (default: 1.8).
• --no_caption: If specified, the script will not generate captions.

Examples

Process all images in a directory:

python -m lora_image_preprocessor.main --input_dir input_images --output_dir output_images

Process only images with faces:

python -m lora_image_preprocessor.main --input_dir input_images --output_dir output_images --face_only

Remove the background from all images:

python -m lora_image_preprocessor.main --input_dir input_images --output_dir output_images --remove_bg

Process images without generating captions:

python -m lora_image_preprocessor.main --input_dir input_images --output_dir output_images --no_caption