LoRAdo

LoRAdo

Prepare datasets for Kohya_ss (a UI for https://github.com/kohya-ss/sd-scripts.)

We plan to integrate training directly into this tool

• Intro
• Why?
• Screenshots
• Features
• Getting Started with LoRAdo
• Update LoRAdo
• Using LoRAdo
• How LoRAdo Works
• Generation examples

Intro

Welcome to LoRAdo. This toolkit is designed to streamline the process of LoRAs image generation. Suitable for both beginners and experienced developers, it simplifies dataset creation. Initially designed with portraits in mind, LoRAdo is adaptable and works well for various subjects. Key features include face detection and user-friendly cropping tools, ensuring a practical balance between user experience and results.

With LoRAdo, creating datasets for image training is straightforward and efficient.

Why?

We wanted to provide an easy way to train LoRAs for different use cases. During research, we decided to build a few small helper functions to help us gather images and crop them as well as create caption files. This UI offers an easy way to create LoRAs without overthinking it.

Screenshots

The user interface	View and configure images
Project view	The prepared folder

Features

• Dataset creation
• Captioning tools
• Create and edit projects
• support for SDXL and SD-1.5
• SDXL or SD-1.5 resolution adjustment
• Multiple image resolutions
• Kohya_ss (UI) config creator
• Regularisation image tools (WIP)

Getting Started with LoRAdo

1. Prerequisites:

   • Ensure you have node.js (v18.5.0) installed. If not, you can download it from Node.js official site. (use a node version manager) to quickly switch to a different version.
   • There is an open issue with tensorflow. Please make sure to install v18 (not above). The recommended version can be found in .nvmrc
   • If there is a problem with tensorflow, please check this issue

2. Clone and Install Dependencies:

   git clone https://github.com/failfa-st/LoRAdo.git
   cd LoRAdo
   npm install

3. Running the App:

   npm run build
   npm run start

   Your app should now be running on http://localhost:3000. Navigate to this URL in your browser to explore the LoRAdo UI.

Update LoRAdo

1. Stop LoRAdo if it's already running

   • CTRL+C to terminate the process in the terminal

2. Update the code and dependencies:

   git fetch origin
   git reset --hard origin/main
   npm install

3. Run the App:

   • Execute the commands as shown under Running the App above

Using LoRAdo

1. Step 1: Begin by uploading your chosen image(s).
2. Step 2: Let LoRAdo's face detection feature scan the image.
   • If a face is detected, the system identifies it as a potential focal point for cropping.
   • If no face is detected, evaluate if the image is suitable for your needs.
3. Step 3: Add captions to your images directly within the app, enhancing the context and meaning of each.
4. Step 4: Choose your cropping method:
   • Opt for minimal cropping by finding the best-fitting resolution based on the original aspect ratio, while still considering the detected focal point.
   • Or, generate 9 SDXL (3 SD-1.5 respectively) compatible crops for each image.
5. Step 5: With a single click on "prepare", transform your curated images into a dataset , compatible to Kohya_ss.
6. Step 6: Navigate to the training folder:
   • Every dataset here comes with its own config.json file.
   • These configurations are primed for integration into Kohya_ss.

How LoRAdo Works

At the heart of LoRAdo is the intent to simplify the complex.

1. Image Optimization: Start with any image. LoRAdo smartly crops it to fit the ideal SDXL resolution. For images with faces, they become the focal point, ensuring the best visual result. Without a face? No worries, we center-crop it perfectly.

2. Multiple Resolutions: Beyond the default, you have the option to generate images in all 9 SDXL resolutions, as an attempt to increase flexibility without compromising on quality.

3. Configuration Made Easy: LoRAdo generates a configuration file tailored for Kohya_ss. Once your project is set, integrating it with Kohya_ss is a breeze. Simply load the config under the LoRA tab, and you're all set to begin training. And if you wish, Kohya_ss offers more granular configuration options.

While our approach is guided by a strong opinion on usability, it still leaves room for customization. As we evolve, expect enhanced configurability. But rest assured, simplicity will always be our guiding star, ensuring that interacting with LoRAdo remains intuitive.

Example images via @anamnesis33

Image	Link
	https://unsplash.com/photos/mqcYKihgfAo
	https://unsplash.com/photos/06TuQM7RSP4
	https://unsplash.com/photos/AUJhl146mBY
	https://unsplash.com/photos/8OWttYqN47I

This tool provides an opinionated configuration and approach to training flexible LoRAs. We are constantly researching to improve the default settings provided by this tool. Advanced (detailed) configuration is planned for future releases.

Generation examples

These examples were generated from a LoRA, trained on a dataset that was prepared with this tool/approach

Feedback and Support: Encountered an issue or have a suggestion? Join our Discord community or open an issue on GitHub. We'd love to hear from you!

Node Version Managers

Node.js has a list of alternative installation options.

Cross Platform

• fnm

Windows

• nvm-windows
• nodist
• nvs

macOS/Linux

• nvm
• n