ImgCap-LoRA

LoRA Dataset Builder

A lightweight, Gradio-based web interface designed to streamline the manual preparation of image datasets for LoRA (Low-Rank Adaptation) and DreamBooth training.

This tool solves the tedious task of resizing images and writing matching .txt caption files simultaneously, ensuring your training data is perfectly formatted for Image Diffusion trainers like Kohya_ss or OneTrainer.

🚀 Features

• ZIP Upload Support: Batch upload your raw images in a single .zip file.
• Dual-Path Saving: Automatically organizes your data into data/images and data/captions directories.
• Intelligent Resizing: Built-in presets for common training resolutions (512px, 768px, 1024px) using high-quality Lanczos or Bicubic interpolation.
• Dynamic Captioning: Automatically generates captions using a base character name and optional "add-on" tags for specific poses or backgrounds.
• Manual Control: Navigate through your dataset with "Next" and "Previous" buttons to ensure every image is curated specifically.

📖 How to Use

1. Enter Character Name: Input the unique trigger word/name for your LoRA (e.g., anshu).
2. Upload ZIP: Upload a zip file containing your raw .jpg, .png, or .webp images.
3. Process Images:

• Select the Body Type (Face, Half, Full) to append to the filename.
• Choose a Resize Preset to ensure uniform dimensions.
• Add Caption Add-ons (e.g., "sitting in a cafe, looking at camera").

4. Save: Click "Save Image + Caption". The tool will save the processed image and a .txt file with the same name into the data/ folder.

📂 Project Structure

├── data/
│   ├── images/      # Processed & resized training images
│   └── captions/    # Matching .txt files for each image
├── temp/            # Temporary extraction folder
└── app.py           # Main Gradio application code