A lightweight Python tool for adding Kohya/A1111-compatible tag frequency metadata to LoRA (Low-Rank Adaptation) safetensors files. This tool fixes missing tag metadata in LoRA files trained with AI-Toolkit or similar programs, ensuring proper display in A1111/Forge/ForgeNeo.
- Two Operating Modes — Dataset mode (with caption files) or Manual mode (trigger words only)
- Automatic Tag Frequency Calculation — Scans caption files and counts tag occurrences across your dataset
- Manual Trigger Word Input — Add metadata without needing the original training dataset
- Native Folder Browser — Easy dataset selection with visual folder picker
- Flexible Path Support — Use subfolders or any custom path on your system
- Kohya/A1111 Compatible — Adds standard metadata fields (
ss_tag_frequency,ss_dataset_dirs, etc.) - Non-Destructive Processing — Creates new files with metadata while preserving originals
- Live Preview & Validation — Real-time feedback on manual tag input
- Smart Dependency Management — Auto-installs required packages in isolated virtual environment
- Portable & User-Friendly — Works from any folder with clear error messages
- Python 3.11+ — Download here
- Windows OS — Batch file automation (Linux support planned)
- .safetensors LoRA file — Your trained model
- Training dataset (optional) — Images with corresponding
.txtcaption files (only needed for Normal Mode)
-
Clone or download this repository:
git clone https://github.com/LindezaBlue/Dataset-Metadata-Injection.git cd Dataset-Metadata-Injection -
No additional setup required! The batch files handle dependency installation automatically.
The easiest way to use this tool is through the visual web interface.
On first run, the program automatically creates two folders:
Model to Repair/— Place your LoRA files hereUpdated LoRA/— Updated files will be saved here
-
Place your LoRA file (
.safetensorsformat) in theModel to Repair/folder -
(Optional) Place your dataset subfolder inside
Model to Repair/
Example:Model to Repair/my_character/- Place all your training images (
.png,.jpg,.jpeg) in this folder - Ensure each image has a matching
.txtcaption file with comma-separated tags
Example:image1.png→image1.txtcontaining1girl, blue hair, smiling, outdoors
OR use a dataset folder anywhere on your system (can browse to it in the UI)
- Place all your training images (
- Place your LoRA file (
.safetensorsformat) in theModel to Repair/folder - That's it! You'll enter trigger words directly in the UI
- Double-click
Run Gradio UI.bat
The batch file will automatically:- Detect your Python installation
- Create a virtual environment (first run only)
- Install required dependencies (gradio, safetensors, torch CPU-only, etc.—first run only)
- Launch the web interface in your default browser at http://127.0.0.1:7860
The interface guides you step-by-step with two modes:
-
Select your LoRA file
- Choose your LoRA Filename from the dropdown
-
Select your dataset
- Choose a Dataset Subfolder from the dropdown, OR
- Click 📁 Select Dataset Folder to browse to any folder, OR
- Paste a custom path directly
-
Scan the dataset
- Click 🔍 Scan Dataset / Parse Tags
- The tool will automatically read all
.txtcaption files and count tag frequencies - Results appear in the Tag Frequencies box with status feedback
-
Inject the metadata
- Click 💾 Inject Metadata & Save
- A new LoRA file with embedded metadata will be created
- Find it in the
Updated LoRA/folder (filename ends with_with_tags.safetensors)
-
Select your LoRA file
- Choose your LoRA Filename from the dropdown
-
Enable manual mode
- Check the "Manual trigger word mode" checkbox
-
Enter trigger words
- Type your trigger words (comma-separated) in the text box
- Set the tag frequency (default: 1)
- Watch the live preview update in real-time
-
Inject the metadata
- When the status shows 🟢 Ready, click 💾 Inject Metadata & Save
- Find the updated LoRA in
Updated LoRA/folder
-
When you're done, just close the browser tab and Terminal/CLI to exit the program.
A launch script run_gradio_ui.sh is provided for Linux systems.
It mirrors the Windows batch file, creating a virtual environment, installing dependencies, and launching the Gradio UI.
Dataset-Metadata-Injection/
├── Run Gradio UI.bat # Launch web interface (Windows)
├── run_gradio_ui.sh # Launch web interface (Linux)
├── gradio_ui.py # Gradio interface code
├── Metadata_Injection.py # Backend code (where the magic happens)
├── requirements.txt # Python dependencies
├── Model to Repair/ # Input folder (auto-created)
│ ├── your_lora.safetensors # Your LoRA files
│ └── [dataset_folders]/ # Optional: dataset subfolders
└── Updated LoRA/ # Output folder (auto-created)
This tool injects the following Kohya SS / A1111-compatible fields:
| Field | Description |
|---|---|
ss_tag_frequency |
JSON object mapping tags to occurrence counts |
ss_dataset_dirs |
Dataset folder names and image counts |
ss_resolution |
Training resolution (default: 1024,1024) |
ss_num_train_images |
Total number of training images |
All dependencies are auto-installed in an isolated virtual environment:
- safetensors — Safe model file loading/saving
- torch (CPU-only) — Minimal PyTorch for tensor operations
- gradio — Web-based UI framework
- packaging — Version management utilities
- Python: 3.11 or later (auto-detected from
%LOCALAPPDATA%\Programs\Python) - Disk Space: ~500MB for virtual environment
- Platform: Windows (Linux support in development)
Problem: Files not found in expected locations
Solutions:
- Verify your LoRA file is in the
Model to Repair/folder - For Normal Mode: verify dataset folder exists (in
Model to Repair/or custom path) - For Manual Mode: no dataset needed, just the LoRA file
Problem: Python installation not detected
Solutions:
- Install Python 3.11+ from python.org
- Use "Install for current user" option during installation
- Script auto-searches
%LOCALAPPDATA%\Programs\Python\
Solution: Delete the venv/ folder and re-run the batch file to recreate it cleanly
Ensure your .txt files use comma-separated tags:
# Correct
1girl, blue hair, smiling, outdoors, school uniform
# Incorrect (will not parse properly)
1girl blue hair smiling outdoors school uniform
-
Implement referenced .json config file for easier editing -
Create Gradio web interface for simplified user experience -
Add manual trigger word mode (no dataset required) -
Add native folder browser for easy path selection -
Support for Linux platforms - Add batch processing for multiple LoRAs
- Tag frequency visualization charts
- Export/import metadata presets
- Any suggestions from the community
Contributions are welcome! Here's how you can help:
- Report bugs via GitHub Issues
- Submit pull requests for new features or fixes
- Share feedback and suggestions for improvement
- Star this repo if you find it useful!
This project is licensed under the
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).
See the LICENSE file for full details.
You are free to use, modify, and share this project, but only for non-commercial purposes, and you must give proper credit to the original author.
If you remix or build upon this work, you must distribute your contributions under the same license!
Created by: LindezaBlue
Special thanks to the AI art community for feedback and testing!
If you find this tool helpful:
- ⭐ Star this repository
- 🐛 Report issues to help improve the tool
- 📢 Share with others who might benefit
- This tool modifies metadata only, not model weights
- Designed for AI-Toolkit trained LoRAs missing standard metadata
- Compatible with Automatic1111, Forge, and ForgeNeo interfaces
- Non-destructive processing ensures your originals remain intact
Questions? Open an issue or check existing discussions!