Sign In
homemodelsimagesvideospostsarticlesbountieschallengeseventsupdatesshop

LoRA Easy Training - Jupyter Remix Edition!

16

ktiseos_nyx's Avatar

ktiseos_nyx

Synthwave Specter

Jul 24, 2025

(Updated: 2 hours ago)

tool guide
byo traininglora training guidegoogle colabtoolslora training
LoRA Easy Training - Jupyter Remix Edition!

LoRA Easy Training - Jupyter Widget Edition

Train LoRAs with guided notebooks instead of confusing command lines

This is a user-friendly LoRA training system based on proven methods from popular Colab notebooks. Instead of typing scary commands, you get helpful widgets that walk you through each step. Works on your own computer or rented GPU servers.

Forked from AndroidXXL and Jelosus2 as well as heavily borrowing from Holostrawberry and Linaqruf's work.  We're submoduled back end with Derrian, SD-Scripts and the KohakuBlueleaf Lycoris system!  (Pray tell the thing works after I broke it again!)

https://github.com/Ktiseos-Nyx/Lora_Easy_Training_Jupyter

MOVED PLEASE. GO HERE: https://github.com/Ktiseos-Nyx/Ktiseos-Nyx-Trainer

Working on adding finetuning and other things to the ecosystem

About

  • Widget-based interface designed for both beginners and advanced users

  • Please note this is STILL a work in progress.

  • Testing was only done on a singular RTX 4090 on a Vast AI Docker Container with pre installed SD WEB UI FORGE.

  • Results MAY vary, please feel free to report issues as you see fit.

  • The system has been recently streamlined with improved widget organization and calculator accuracy.

✨ What You Get

🎓 Beginner-Friendly

  • Helpful explanations for every setting (no more guessing!)

  • Step calculator shows you exactly how long training will take

  • Warnings when settings don't work together

🧪 Advanced & Amazing Standard Options

  • Memory-efficient optimizers (CAME, Prodigy Plus)

  • Special LoRA types (DoRA, LoKr, LoHa, IA³, BOFT, GLoRA)

  • Memory-saving options for smaller GPUs

🛠️ Easy Setup

  • Two simple notebooks: one for datasets, one for training (Third snuck in for utilities)

  • Works with VastAI and other GPU rental services

  • Checks your system automatically

📊 Dataset Tools

  • Auto-tag your images (WD14 for anime, BLIP for photos)

  • Add/remove tags easily

  • Upload ZIP files or folders

🚀 Quick Start

What You Need

  • Computer: Windows, macOS, or Linux

  • Python: Version 3.10 or newer

  • GPU: NVIDIA GPU with 8GB+ VRAM recommended (can work with less)

  • Git: For downloading this project (explained below)

Installation

  1. Get Git (if you don't have it)

    Git is a tool for downloading code projects. Don't worry - you just need to install it once and you're done!

    Check if you already have Git: Open your terminal/command prompt and type git --version. If you see a version number, you're good to go!

    If you need to install Git:

    • Windows: Download from git-scm.com and run the installer

    • Mac: Open Terminal and type xcode-select --install

    • Linux: Type sudo apt-get install git (Ubuntu/Debian) or use your system's package manager

  2. Download This Project

    Open your terminal/command prompt and navigate to where you want the project folder. Then run:

    git clone https://github.com/Ktiseos-Nyx/Lora_Easy_Training_Jupyter.git
cd Lora_Easy_Training_Jupyter

  3. Run Setup

    This automatically installs everything you need:

    Mac/Linux:

    chmod +x ./jupyter.sh
./jupyter.sh

    Windows (or if the above doesn't work):

    python ./installer.py

    Just wait for it to finish - it downloads the training tools and sets everything up.

Start Training

If using VastAI or similar: Jupyter is probably already running - just open the notebooks in your browser.

If on your own computer: Start Jupyter like this:

jupyter notebook

Then open these notebooks:

  1. Dataset_Maker_Widget.ipynb - Prepare your images and captions

  2. Lora_Trainer_Widget.ipynb - Set up and run training

  3. LoRA_Calculator_Widget.ipynb - Calculate training steps (optional)

📖 How to Use

Step 1: Prepare Your Images

Open Dataset_Maker_Widget.ipynb and run the first cell:

# This starts the dataset preparation tool
from widgets.dataset_widget import DatasetWidget
dataset_widget = DatasetWidget()
dataset_widget.display()

Upload your images (ZIP files work great!) and the system will auto-tag them for you.

(This part is actually pretty much the same as the notebooks I was inspired from, and will not lie to you I just had Claude and Gemini reformat it so the wording fit what we're doing in Jupyter, it's either Android XXL or Jelosus2's words originally) 

To use custom models or VAEs, you need to provide a direct download link. Here’s how to find them on popular platforms:

From Civitai

Method 1: Using the Model Version ID

This is the easiest method if a model has multiple versions.

  1. Navigate to the model or VAE page.

  2. Look at the URL in your browser's address bar. If it includes ?modelVersionId=XXXXXX, you can copy the entire URL and paste it directly into the widget.

  3. If you don't see this ID, try switching to a different version of the model and then back to your desired version. The ID should then appear in the URL.

How to get a link from Civitai using the version ID

Method 2: Copying the Download Link

Use this method if the model has only one version or if a version has multiple files.

  1. On the model or VAE page, scroll down to the "Files" section.

  2. Right-click the Download button for the file you want.

  3. Select "Copy Link Address" (or similar text) from the context menu.

How to get a link from Civitai by copying the download address

From Hugging Face

Method 1: Using the Repository URL

  1. Go to the main page of the model or VAE repository you want to use.

  2. Copy the URL directly from your browser's address bar.

How to get a link from Hugging Face using the repository URL

Method 2: Copying the Direct File Link

  1. Navigate to the "Files and versions" tab of the repository.

  2. Find the specific file you want to download.

  3. Click the "..." menu to the right of the file size, then right-click the "Download" link and copy the link address.

How to get a link from Hugging Face by copying the direct file address

Step 2: Train Your LoRA

Open Lora_Trainer_Widget.ipynb and run the cells to start training:

# First, set up your environment
from widgets.setup_widget import SetupWidget
setup_widget = SetupWidget()
setup_widget.display()

# Then configure training
from widgets.training_widget import TrainingWidget  
training_widget = TrainingWidget()
training_widget.display()

3. Extras

🧮 Quick Training Calculator

Not sure about your dataset size or settings? Use our personal calculator:

python3 personal_lora_calculator.py

This tool helps you:

  • Calculate optimal repeats and epochs for your dataset size

  • Get personalized learning rate recommendations

  • Estimate total training steps

  • Build confidence for any dataset size (no more guesswork!) 🎯

🔧 Architecture

Core Components

  • core/managers.py: SetupManager, ModelManager for environment setup

  • core/dataset_manager.py: Dataset processing and image tagging

  • core/training_manager.py: Hybrid training manager with advanced features

  • core/utilities_manager.py: Post-training utilities and optimization

Widget Interface

  • widgets/setup_widget.py: Environment setup and model downloads

  • widgets/dataset_widget.py: Dataset preparation interface

  • widgets/training_widget.py: Training configuration with advanced mode

  • widgets/utilities_widget.py: Post-training tools

🐛 Troubleshooting

Support

  • GitHub Issues: Report bugs and feature requests

  • Documentation: Check tooltips and explanations in widgets

  • Community: Share your LoRAs and experiences!

🏆 Credits

This project is built on the work of many awesome people:

Training Methods:

Notebook Inspirations:

  • AndroidXXL, Jelosus2, Linaqruf - Original Colab notebooks that made LoRA training accessible

Community:

"Either gonna work or blow up!" - Made with curiosity! 😄

🔒 Security

Found a security issue? Check our Security Policy for responsible disclosure guidelines.

📄 License

MIT License - Feel free to use, modify, and distribute. See LICENSE for details.

🤝 Contributing

We welcome contributions! Check out our Contributing Guide for details on how to get involved. Feel free to open issues or submit pull requests on GitHub.

Made with ❤️ by the community, for the community

FAQ

  • Q: "Will this work on Windows?"

  • A: No clue. Feel free to test and Fork away!

  • Q: "Will this upload directly to Civitai"

  • A: No.

  • Q: "Will this train SD3 and Flux?"

  • A: Not YET, the widgets haven't been enabled yet for this.

  • Q: "Does it REQUIRE VAST AI?"

  • A: No, and i'm kinda cranky Claude wrote the damn Readme like it's the only system it'll work on.

  • Q: "UR NOT A PROGRAMMERR111oeneoneone!!!"

  • A: Did I say i was? :D You already know I whip things together. by the seat of my pants like a crazed ADHD Squirrel with LLMs. You're lucky I remember how these things work in the back end, :D

  • Q: "Dusk, why did you reinvent the wheel" - I'm sure my peers are sayin' this XD

  • A: Because i'm one of few that misses doing this, and I hate how Colab is stupidly janky, Holo's work is why I learned to do Loras so well. I promised I was going to move off Civitai for lora training and learn things, but all the systematic approaches with a ll the new options just didn't work at the moment. So I forked Jelosus and Android XL's notebooks, Forked Holo's - did a frantic post norovirus debug & rejigging session and gave you this!

  • Q: "Will it work on a Mac?"

  • A: ...... Sell your mac, get a PC. - Sincerely a Mac user. xD True Answer: ARM mac i'm not sure, I do'nt know how Cuda effects that I am still on an intel mac.

  • Q:"Will it work on less than a 4090"

  • A: At the moment, i'm NOT SURE because I haven't figured out how to get Claude to pay attention to direct directions without it forgetting things for 30 seconds. Both Gemini and Claude were playing fake data games with me lol. Once I get the hang of everything that the other notebooks have, and the more data I get my paws on for education and training details the more this project will shine better.

  • Q: "Why not just use Civitai"

  • A: I'm not telling you not to! Civitai's lora trainer was the backbone for MANY people and it's simple, easy and "AFFORDABLE" yes - but I'm an idiot and I like to waste buzz, and this option is acutally cheaper on my buzz balance.

Future Focusing:

Inference * FiftyoneAI were coming and still are eventually but give me time i'm still learning how all this stuff works. 

We're still ironing out bugs on main, so the repo will tell you to use unified. :) 

16