This article explains the important steps of tagging and trigger words in LoRA training for beginners. There is also an article that covers more advanced topics, so once you understand this content, be sure to check those out as well.

1. What is LoRA?

LoRA (short for Low-Rank Adaptation) is a convenient tool that can be used when generating images with Stable Diffusion. Simply put, it learns specific characters, compositions, or styles from a small amount of image data and applies that knowledge to image generation.

Difference from a Checkpoint:
- Checkpoint: A large model that can generate images on its own (trained on massive amounts of data).
- LoRA: A small model used in addition to a Checkpoint (can learn effectively from limited data). It is created by additional training on a Checkpoint.

For example, if you create a LoRA that learns the features of your original character or pet, you can generate images of that character or pet in various scenes and compositions! As a general rule, a LoRA should focus on learning one element only.

2. What is Tagging?

Tagging is the process of describing “what elements are included in an image” using words. In this article, we’ll focus on Danbooru-style prompt tags. Think of them as equivalent to the prompts (commands) used during image generation.

For example, for a photo of a cat, tags might look like this:
1 cat, aqua eyes, calico cat, depth of field, looking up, photo realistic, street, walking

3. What is a Trigger Word?

A trigger word is like a keyword or “password” that activates the effect of the LoRA you created. When you include this word during image generation, the features you trained will be reflected in the output. Just as the name suggests, it acts as a “trigger.”

Tips for setting a trigger word:
- Use a word that does not overlap with existing Danbooru tags.
- Choose a word that intuitively reflects what was trained.
- Keep it relatively short.

For instance, using the cat example above, if you trained the features of a calico cat (1 cat, aqua eyes, calico cat), you might set the trigger word as something like my_cat.

4. Practical Tagging for LoRA Training

Here are the steps for adding a trigger word and editing tags in a dataset for LoRA training:
- Add the trigger word to all training images.
- Remove the original tags that describe features now represented by the trigger word.
- Keep tags that describe composition, background, or other elements not being trained.

Let's compare the original tags and the training tags using the cat example:

Original tags:
1 cat, aqua eyes, calico cat, depth of field, looking up, photo realistic, street, walking

Training tags:
my_cat, depth of field, looking up, photo realistic, street, walking

By doing this, the AI learns to associate the trigger word ‘my_cat’ with the features of a calico cat. You might feel uneasy about removing tags that describe the subject, but if you leave them in, the AI may interpret my_cat and 1 cat, aqua eyes, calico cat as different concepts. Put simply, you need to “absorb” the target features into the trigger word among the uniformly applied tags. Tags that describe elements not targeted for training, like composition or background, should remain.

5. Summary

This was a brief explanation of tags and trigger words. The world of LoRA is deep, and by experimenting and gaining experience, you’ll gradually get closer to generating the images you envision. Understand the basics of trigger words and tags, and try creating your own original LoRA!