LoraTraining — Complete Krea 2 LoRA Dataset Manager, Training Queue & ComfyUI Validation Suite

LoraTraining was developed to simplify and streamline the entire process of preparing, managing, training, and validating Krea 2 LoRAs. In fact, every single one of my LoRAs here were directly made using this tool which I decided to share with the community. I'm not going to claim this is the best tool that exists, but that's the one I built and I use... and it does exactly what it says: everything I need from start to finish.

What started as a personal tool to organise training datasets has evolved into a comprehensive Windows desktop application for dataset management, AI-assisted captioning, image editing, training configuration, automated job queues, and advanced ComfyUI validation.

It is built on top of AI Toolkit for the actual LoRA training. It does not replace or reimplement the trainer. Instead, it provides a comprehensive interface to prepare datasets, configure jobs, automate training sessions, monitor progress, and evaluate results.

⭐ PLEASE give the project a STAR on GitHub if you find it useful — it would be greatly appreciated!

Proven through everyday use: LoraTraining isn't just an experimental interface. I've been using it as my primary LoRA training management tool for months and have successfully trained over 450 LoRAs with it.

Whether you're training your first few LoRAs or managing dozens of datasets and hundreds of checkpoints, LoraTraining helps you spend less time managing files and scripts — and more time improving your models.

Instead of juggling countless folders, caption files, YAML configurations, training scripts, and generated checkpoints, LoraTraining brings everything together into a single, organised workspace.

The initial setup requires a working AI Toolkit and (optional) ComfyUI installations (for validation), but once configured, LoraTraining provides a streamlined workflow for managing datasets, automating training queues, and evaluating checkpoints.

If you have any questions or need additional help, read the DOCUMENTATION here or even better, you can provide your favourite AI assistant (ChatGPT, Claude, Grok, Gemini, Qwen, etc.) with the complete documentation below and ask practically anything about the application: AI SUPPORT FILE

The documentation is specifically designed for AI-assisted support, including detailed instructions, configuration references, troubleshooting, and a comprehensive FAQ.

✨ Main Features

🗂️ Dataset & Library Management

• Organise LoRA training datasets into categories and folders.

• Discover and import existing image datasets.

• Browse images through an integrated thumbnail grid.

• Manage multiple training datasets from a single workspace.

• Track each dataset's progress through its lifecycle: Ready → Trained → Validated → Done.

• Duplicate, rename, reorganise, and manage datasets.

• Automatically scan image folders and identify relevant training files.

• Detect missing captions and other dataset issues.

• Efficiently manage large collections with progress reporting and cached previews.

📝 AI Captioning & Caption Management

• Integrate JoyCaption for automated image captioning using a local AI model.

• Generate captions for individual images or process images in batches.

• Edit captions directly within the application.

• Apply bulk caption operations across selected images.

• Add identity tags and manage trigger words.

• Choose whether to caption only missing files, append to existing captions, or regenerate them.

• Use caption guidance to influence automated descriptions.

• Validate captions before training.

JoyCaption runs through an external Python environment using the bundled captioning script. LoraTraining manages the process and resulting caption files.

🖼️ Integrated Image Management & Editing

• Browse, inspect, and organise training images without leaving the application.

• Edit images using an integrated image editor with all the most commonly needed editing options, including Content-Aware removal using ONNX AI model.

• Move or copy selected images between datasets.

• Remove unwanted images and manage associated caption files.

• Select dataset cover images.

• Open images in Photoshop for external editing.

• Optionally enable NTFS image version history to preserve previous edits.

• Keep image files and captions organised within ordinary folders on disk.

🌐 Integrated Image Scraper

• Browse Pinterest and Google Images from within the application.

• Collect reference and training images directly into the selected dataset.

• Reduce the need to manually download, rename, and move images between folders.

• Review and organise collected images before training.

⚙️ LoRA Training Configuration

• Create and manage training jobs through a dedicated interface.

• Configure essential parameters such as:

  • Training steps and checkpoint save intervals.

  • LoRA rank.

  • Learning rate.

  • Training resolution.

  • Dataset paths and trigger words.

  • LoRA output names.

• Maintain reusable default training configurations.

• Configure the Krea 2 base checkpoint and architecture globally.

• Access Low VRAM, model quantisation, and transformer layer-offloading controls.

• Generate the YAML configuration consumed by AI Toolkit.

• Inspect and edit job configurations without manually creating training scripts.

The actual training is performed by AI Toolkit. LoraTraining manages the configuration, process execution, and results.

🚀 Automated Training Queue

One of LoraTraining's main objectives is to make training multiple LoRAs considerably easier.

• Create and enqueue multiple training jobs.

• Run jobs sequentially using a FIFO training queue.

• Reorder queued jobs.

• Monitor training progress and live logs.

• View estimated durations and completion times.

• Pause, stop, restart, and retry jobs.

• Automatically continue to the next job after failures by default.

• Introduce a cooldown period between training jobs.

• Prevent Windows from entering sleep mode during active training.

• Apply process-priority management while training is running.

• Keep job and queue states organised across application sessions.

Prepare multiple datasets, queue the jobs, and let LoraTraining orchestrate the training sessions.

Only one training job runs at a time, avoiding simultaneous training processes competing for GPU memory.

🛡️ Windows Training Reliability

LoraTraining includes additional process-management functionality aimed at improving the reliability of long-running Krea 2 training sessions on Windows.

• Bundled training bootstrap for AI Toolkit.

• A safetensors loading workaround targeting Windows OSError 1455 memory-commit failures.

• Training process-tree management.

• GPU cooldown between jobs.

• Live progress and failure reporting.

• Diagnostic logs for investigating failed training sessions.

These features improve orchestration and address specific Windows problems, but do not replace AI Toolkit's own environment requirements or troubleshooting.

🎨 Advanced ComfyUI LoRA Validation

Training a LoRA is only part of the process. Evaluating the results is equally important.

LoraTraining integrates with ComfyUI to help compare checkpoints, assess training progress, and generate validation images.

Validation Steps

• Generate preview images from different saved training checkpoints.

• Compare how the LoRA evolves throughout training.

• Identify checkpoints that produce the most desirable results.

Validation Weight

• Test different LoRA strengths.

• Compare how strongly the trained concept affects the generated images.

• Evaluate checkpoints using consistent test configurations.

Model Stress-Testing

• Test LoRAs against multiple configured validation models.

• Compare results and identify potential weaknesses.

• Continue through test sets even when individual models fail.

Additional tools

• Stress-test comparison.

• Image upscaling.

• Context diversification.

• Expression diversification.

• Preview ratings and checkpoint evaluation.

• Reusable validation sets and configurable ComfyUI workflows.

• Wildcard-aware prompt editing for supported preview tools.

The repository includes LoraTraining custom nodes and starter workflows for ComfyUI integration.

ComfyUI itself, additional required custom-node packs, and model weights must be installed separately.

📊 Checkpoint & Training Lifecycle Management

• Browse LoRA checkpoints generated during training.

• Track completed and pending training steps.

• Organise validation previews.

• Keep selected checkpoints available for ComfyUI testing.

• Manage additional training sessions.

• Track datasets through Ready, Trained, Validated, and Done states.

• Associate validation results with the corresponding dataset and checkpoints.

This makes it easier to compare different stages of training without manually organising hundreds of generated files.

💾 Local Library & Data Management

• Keep datasets, captions, jobs, configurations, and application workspace data in a library folder of your choice.

• Store images as ordinary files rather than inside a proprietary database.

• Maintain ComfyUI workflows, thumbnails, and job metadata alongside the library.

• Export and import application settings.

• Back up the library by copying its files and folders.

• Optionally use NTFS alternate data streams for image version history.

Your training images and outputs remain on your own storage.

🔄 Application Updates

• Check for new releases directly from the application.

• Download installers from GitHub Releases.

• Read the What's new release notes.

• Install updates without manually searching the repository.

• Optionally use locally built installers from a configured release folder.

💻 Desktop Application

LoraTraining is a standalone Windows desktop application developed with Electron, React, and TypeScript.

• Platform: Windows 10 / 11

• Supported training architecture: Krea 2 text-to-image

• Training backend: AI Toolkit

• Automated captioning: JoyCaption

• Validation and image workflows: ComfyUI

• Installation: Windows executable installer

LoraTraining is NOT a standalone training engine and is NOT a ComfyUI custom node.

It is a desktop management and orchestration application that integrates with existing installations of those tools.

Dataset organisation, manual caption editing, and other library-management features remain usable without configuring every external tool.

🔧 External Tools & Requirements

Important: LoraTraining is intended for experienced local-AI users who already have their training and generation environments working.

• AI Toolkit: Performs the actual Krea 2 LoRA training

• JoyCaption model: Automated image captioning through the bundled Python script

• ComfyUI: Checkpoint validation, image generation, upscaling, and diversification

LoraTraining provides the interface for configuring and orchestrating these tools, but does not install or maintain their Python environments, CUDA dependencies, model weights, or third-party custom nodes.

Users should first verify that AI Toolkit can successfully train Krea 2 LoRAs independently. ComfyUI must also be functional independently if validation or preview features are required.

Complete LoraTraining setup instructions

🔗 Downloads & Source Code

GitHub repository:
https://github.com/aipower66669/LoraTraining

Latest Windows installer:
https://github.com/aipower66669/LoraTraining/releases/latest

Complete documentation and FAQ:
https://github.com/aipower66669/LoraTraining/blob/main/docs/all-in-one.md

User guide:
https://github.com/aipower66669/LoraTraining/blob/main/docs/USER_GUIDE.md

The repository includes source code, documentation, screenshots, installation instructions, ComfyUI integration nodes, and starter workflows.

🆕 What's New in Version 0.5

Image editing

• Text annotations keep the rectangle you drag, with the same move and resize handles as a rectangle. Double-click to type.

• Undo last drops only the newest NTFS version. Revert to original still clears the whole stack.

• Edit in Photoshop asks whether to open the original or commit the latest edit when that image has a version stack.

• Saving an edit of a .jfif writes a JPEG.

Version 0.4

Library management and workflow improvements

• Optional NTFS image history: Image versioning is now opt-in. With the option disabled, image edits overwrite files normally.

• Capability-based interface: Features that require JoyCaption, AI Toolkit, or ComfyUI appear when the corresponding tool paths are configured. Manual dataset management remains available.

• Integrated updates: Check GitHub Releases, view release notes, and install newer versions directly from the application.

Performance and training reliability

Building on improvements introduced in v0.3.0:

• More efficient dataset discovery with visible progress reporting.

• Shallow checkpoint scanning and cached validation previews.

• Faster batch enqueue operations.

• Improved queue status synchronisation.

• Windows safetensors bootstrap addressing OSError 1455.

• Process management and cooldowns for overnight training queues.

Documentation and support

• Comprehensive installation and configuration guides.

• Detailed feature documentation.

• Extensive FAQ and symptom-based troubleshooting.

• A consolidated all-in-one.md specifically designed for AI-assisted support.

• Explicit separation between LoraTraining functionality and the responsibilities of AI Toolkit, JoyCaption, and ComfyUI.

Full release notes:

https://github.com/aipower66669/LoraTraining/blob/main/RELEASE_NOTES.md

📜 License

LoraTraining is free for noncommercial use under the PolyForm Noncommercial License 1.0.0.

Commercial use of the application requires separate permission from the copyright holder.

Full license:
https://github.com/aipower66669/LoraTraining/blob/main/LICENSE.md

Built for people who take LoRA training seriously.

Spend less time organising datasets, editing captions, preparing YAML files, monitoring training scripts, and comparing checkpoints — and more time training better LoRAs.