🪄LoRA Training

A. Basics of LoRA Training

This guide will walk you through the step-by-step process of training a LoRA/Spell on Moescape AI The first section covers the basics, while the second dives into advanced techniques to help you create even better LoRA/Spells.

Below are the default prices for LoRA training (subject to change).

I. The First Page of Lora Training

LoRA Training Interface Overview

When you click "Train a LoRA", you'll see the front page with numbered sections, each serving a specific purpose. Here's what each section does:

1. Name

Give your LoRA/Spell a unique name. It can represent a character, style, object, or background. A distinct name helps differentiate it from other LoRAs/Spells.

2. Trigger Words

These are the keywords that link to the images being trained, triggering the desired effect. For character LoRAs, use a unique trigger word to avoid conflicts with existing models. For example, instead of just Sailor Moon, use 'TrigSailorMoon' to ensure precise results. The same applies to artist tags.

3. Description

Provide a clear explanation of what your LoRA/Spell does. This helps users understand its purpose and functionality.

4. Base Model

Select the model type you want to train your LoRA/Spell on. It uses a default model to integrate your images properly. For example, if you're training a LoRA for the Illustrious model, select it here.

• You'll see an "Advanced Options" toggle below—keep it off unless you're experienced. Advanced settings will be covered in the next tutorial later at the end of this page.

5. Custom Model

If you're experienced, you can select a different model from the Model Hub for training. However, beginners should stick with the base model for better stability.

6. Image Style + Terms & Conditions

Choose the visual style of your LoRA/Spell: Anime or Realistic.

This setup ensures your LoRA/Spell is trained effectively with the right parameters!

Once you've completed all the required fields and accepted the terms and conditions, click "Next" to proceed to the most crucial step—creating your dataset!

II. Making your Dataset of Images

Now, we move on to the second step of creating a basic LoRA/Spell—gathering and organizing your training dataset. The number of images required varies depending on the type of LoRA/Spell, whether it's for a character, art style, background, or concept (e.g., objects, clothing, poses). First, let's explore different sites that use tagging systems to help you source the right images for training.

1. https://danbooru.donmai.us - Danbooru Image Hoster

2. https://gelbooru.com - Gelbooru Image Hoster

3. https://e621.net - E621 Image Hoster

These are just a few sites where you can search for the images you need.

Each LoRA/Spell type requires a specific number of images for proper training:

• Character LoRA (20–30 images) – Include a variety of camera angles: front, back, side, above, and below.

• Style LoRA (Artist, Background Artist) – No set limit, but ensure you have enough images that clearly showcase the artist’s style. The only constraint is the platform’s upload limit.

• Concept LoRA (Clothing, Objects, Poses, etc.) – Requires 30–50 images. Capture multiple angles to accurately train the model for the desired concept.

Make sure your dataset is diverse and well-structured for the best results!

Uploading Your Dataset & Starting Training

The image above shows the next step—uploading your dataset images. Currently, there is a 100-image limit for uploads, though this may change in the future. In the upper right corner, you’ll find the Auto-Tag option (circled in red), which scans and automatically tags your images.

Follow these steps to upload and begin training your LoRA/Spell:

Step 1:

Click "Drag and Drop" or "Click to Select" to upload your images (max 100).

Step 2:

After uploading, click "Auto Label" to automatically generate tags for each image.

Step 3:

Manually add your chosen trigger word to each image’s tags to ensure proper activation.

Step 4:

Review and refine your tags:

• Character LoRA: Remove irrelevant tags like 1girl, solo, or white background.

• Style LoRA: Keep all tags, as they contribute to the stylistic training.

• Concept LoRA: Follow the same tag-cleaning rules as Character LoRAs.

Step 5:

Once satisfied with your dataset, scroll to the bottom and click "Start" to begin training. The page will display the credit cost, and training typically takes 1–3 hours. After completion, you can publish your LoRA/Spell in the Model Hub.

B. Advanced Steps of LoRA Training

I. Custom Models

The first advanced setting for LoRA training on Moescape AI is the Custom Model option, labeled "+ Custom Model" in the image above. Clicking this opens the Model Hub, where you can select a custom version of your chosen model. For example, if you're training with Illustrious, you can choose a custom Illustrious model instead of the default IllustriousXL 1.0 model.

II. Advanced Settings

As you can see from above this is the first part of the advanced settings option that is available after you select your model type.

**WARNING** DO NOT MESS WITH THESE OPTIONS UNLESS YOU ABSOLUTELY KNOW WHAT YOU ARE DOING

Advanced Settings for LoRA Training on Moescape AI

For most users, it's best to stick with default settings when enabling Advanced Settings in LoRA training. However, a few key adjustments can improve results:

• Resolution: When training Illustrious LoRA/Spells, it's recommended to use 1024 resolution instead of the default 512 resolution for better quality.

• Keep Tokens: If you choose to shuffle tags, set keep_tokens to 1 to maintain consistency.

Beyond these, avoid modifying other settings unless you have prior experience with LoRA/Spell training. A more detailed tutorial will cover each setting in depth, but for now, only adjust the following key parameters if you know what you're doing:

Key Advanced Settings

1. Unet_lr (Learning Rate)

   1. Controls how fast the model learns. Think of it as the pace at which the model learns from the data.

   2. Too high = overshooting the optimal solution.

   3. Too low = slow training or getting stuck in a suboptimal state.

   4. Optimal range: 0.0001 – 0.0004 (varies by model type).

      1. SDXL/Illustrious/1.5: 0.0004

      2. Pony: 0.0003

2. Text_encoder_lr

   1. Sets the learning rate for the text encoder, which affects the entire U-Net.

   2. Should always be lower than the Unet_lr.

   3. Optimal range: 0.00001 – 0.00005

      1. SDXL/Illustrious/1.5: 0.00003

      2. Pony: 0.00002

3. Network_Dim (Network Dimension) & Network Alpha

   1. Directly related to LoRA training file size and efficiency.

   2. Network Rank: 4–32 (higher values create larger files, but 32 is usually sufficient).

   3. Network Alpha: Should be half of the Network Dimension.

      1. Example: If Network Dim = 32, then Network Alpha = 16.

Final Notes

Other advanced settings should not be adjusted unless you fully understand their impact. This concludes the advanced steps of LoRA training on Moescape AI.

Special thanks to Moescape community member J Ramsey for helping put together this LoRA training guide.