Earngenix Logo
Skip to main content

Glossary · LoRA Training

What Is Captioning in LoRA Training?

By Earngenix Team ·

⚡ Quick Answer

Captioning is writing a short text description for each image in your LoRA training dataset. It tells the training process exactly what's in the picture, so the LoRA learns to connect the right words to the right visual details instead of guessing.

A simple rule of thumb: describe anything you want to be able to change later with your prompt, and leave out anything that should always be present in the LoRA's output.

Where You'll See It

Each image gets its own .txt file with a matching filename, sitting right next to it in the training dataset folder. If the image is portrait_01.png, its caption lives in portrait_01.txt.

Quick Example

A caption file might read: "a woman with red hair, studio lighting, blue jacket, smiling." Later, you could prompt for the same subject in a green jacket outdoors, since jacket color and lighting were described — not baked in as fixed.

Consistency across captions matters more than perfect wording. Using the same style and level of detail across every caption in your dataset trains a more predictable LoRA than mixing very short and very detailed captions.

Frequently Asked Questions

No — leave those out. Anything you leave undescribed but present in every image (like a character’s face) gets baked permanently into the LoRA. Anything you do describe (like "red jacket") stays changeable later through your prompt.

For a subject-focused LoRA, yes — including the trigger word consistently is what lets you reliably call up that subject later. For a pure style LoRA, a trigger word is optional.

Yes, many training tools include automatic captioning using vision models, which can save significant time on large datasets. Reviewing and correcting a sample of the auto-generated captions is still worthwhile before training.

See It In Action

Ready to train your first LoRA?

Our LoRA guide covers using, and getting started with, LoRAs in ComfyUI.

Published: 2026-09-17 · Last updated: 2026-09-17

Discussion

Join the discussion

Sign in to leave a comment or reply

💬

No comments yet

Be the first to share your thoughts!