This tool allows you to merge two LoRA models using adaptive merging strategies. You can choose between a single weighted merge or a mix of three different weight ratios. This utility will guide you step-by-step through the process of selecting and merging your LoRA models.
YouTube on My Merged LoRAs: https://www.youtube.com/watch?v=VUV6bzml2SU
- Adaptive Merging: Merge two LoRA models with adaptive merge strategies.
- Manual Weight Merging: Support for traditional weighted merging, allowing users to force specific weights.
- NEW! Additive Merging: Use 100% of one LoRA and add a specific percentage of another, perfect for enhancing similar concepts.
- Mix or Weighted Options: Choose from single weighted, or create mixed versions for 25%, 50%, and 75% weights automatically.
- User-Friendly Guidance: Easy-to-follow prompts guide you through the setup.
Adaptive merging is a unique approach that adjusts the contribution of each LoRA model dynamically based on their characteristics rather than using fixed or arbitrary weights. This is particularly useful when combining models that have different focuses, such as content and style. If you're interested in the implementation, you can look at the adaptive_merge() function in merge_lora.py.
I currently have over 20 concepts for RPGv6 and need to add around 50 more. I aim to avoid training individual LoRA files for each RPG class and race. During training, I noticed I was losing some ability to render my concepts in different styles (pencil, sketch, real life, concept art) because the initial art style used for training concepts tends to dominate. Adding more diverse images to the dataset didn't fully resolve this.
So, I trained a second LoRA focused on rendering styles of the same classes with specific emphasis on keywords. Instead of merging all of this in a single training session, I use adaptive merging to enhance the power of the style LoRA while balancing my concept LoRA. The improvements were significant enough to make this utility publicly available, hoping it might help others facing similar challenges.
In the future, I might need to train an entire checkpoint, but for now, adaptive merging pushes the limits of my RPG v6 LoRA effectively.
In adaptive merge, the weight value is used when adaptive merging identifies conflicting layers or important overlaps that need prioritization. Sometimes, you may want to force a weight and ignore adaptive merge recommendations. LoRA will behave differently, so having the traditional weight merge approach is valuable for experimentation.
The Additive Merge strategy allows you to use 100% of the first LoRA model while adding a specified percentage of the second model. This method effectively strengthens the influence of the main model while incorporating the desired effects of the secondary model, making it ideal for combining similar subjects trained on different datasets. The Additive Merge always retains the full influence of the main LoRA, then adds the second LoRA at the specified percentage level, enabling a unique way of blending without losing the original strengths of the main model. This approach is perfect for situations where two models cover the same subject but differ in stylistic elements, allowing you to enhance a concept without retraining from scratch.
Special thanks to @Tatzes for inspiring this feature!
- Python 3.6 or higher installed on your system.
- Git installed on your system.
- Basic understanding of using command line or terminal.
-
Click on the Code button at the top right of this page.
-
Copy the URL and open your terminal (Command Prompt on Windows, Terminal on macOS/Linux).
-
Clone the repository with the following commands:
git clone https://github.com/Anashel-RPG/anashel-utils.git cd anashel-utils
For macOS/Linux users, you need to make the start_mac-linux.sh script executable:
chmod +x start_mac-linux.shFor macOS/Linux Users:
Run the following command to start the utility:
./start_mac-linux.shFor Windows Users:
Click on start_windows.bat in the project folder, or run the following command in Command Prompt:
start_windows.batOnce the utility is running, it will guide you through each step. Below is a sample console output to give you an idea of what to expect:
Boot routine initiated...
┌──────────────────────────────── LoRA Merger ────────────────────────────────┐
│ Welcome to the Anashel's LoRA Merging Utility! │
│ │
│ This tool allows you to merge two LoRA models using adaptive merging │
│ strategies. You can choose between a single weighted merge or a mix of │
│ three different weight ratios. The process will guide you through selecting │
│ your LoRA models and merge settings. │
└─────────────────────────────────────────────────────────────────────────────┘
Press Enter to continue...
This utility allows you to set up a merge of LoRA models by selecting two
models and adjusting the merge weight percentage.
Loading LoRA models: 100%|██████████| 2/2 [00:00<00:00, 12.67file/s]
+-------+------------+------------------+-----------+
| Index | LoRA Model | Number of Layers | File Size |
+-------+------------+------------------+-----------+
| 1 | 067-15000 | 380 | 142.55 MB |
| 2 | 071-12000 | 380 | 142.55 MB |
+-------+------------+------------------+-----------+
Select the main LoRA source (1-2): 1
Select the LoRA to merge with (1-2): 2
Enter the percentage to keep from the main model (0-100)
You can also type 'mix' for 25%, 50%, 75% versions: 30
You have chosen to merge:
Main LoRA: 067-15000
Merge LoRA: 071-12000
Merge Strategy: Weighted
Weight Percentage: 30.0%
Is this satisfactory? (no to adjust, yes to continue): yes
Merging completed! ✅
Your new LoRA will start with mrg_ to identify this was the result of a merge, then you will have the first and second LoRA names with a 3-letter tag in the middle. A stands for Adaptive, M stands for Manual, followed by the weight percentage. For example, A25 means that this merge is the result of an Adaptive approach with a 25% weight, while M75 is a manual imposed 75% merge for all layers in the LoRA.
- Python Not Installed: Ensure Python 3.6 or higher is installed on your system. You can download it from python.org.
- Permission Denied (macOS/Linux): Make sure to run
chmod +x start_mac-linux.shto make the script executable. - Dependencies Not Installed: The script will automatically install dependencies listed in
requirements.txt. Ensure you have an active internet connection when running the script.
More LoRA merge options + I plan to upload all the utilities I've coded to assist in my training of RPG v6, including image auto-captioning and style transfer to diversify and enrich datasets. In the meantime, feel free to join my Discord community for updates and discussions: Join my Discord.
You can also download my LoRA: RPG v6 for Flux1 on CivitAI.