A LoRA training dataset example

An image-and-caption dataset pairs each training image with a text file describing it. This guide shows how to organize those pairs, review their contents and export them. Follow your training tool’s documentation for the final folder structure and settings.

Curaset documentation · Updated 2026-10-05

Match image and caption filenames

The image and its caption must have the same base filename: mug_01.jpg goes with mug_01.txt. Use an image format your training tool supports. Each image needs its own caption file; a single TXT file cannot provide separate captions for a whole folder.

dataset/
  mug_01.jpg
  mug_01.txt
  mug_02.jpg
  mug_02.txt

Write a caption for each image

The example below describes a ceramic mug on a wooden table. Compare it with the image and describe what you can actually see. Check your model’s documentation when choosing a caption format, then adapt the example to your own dataset.

A ceramic mug on a wooden table, soft window light, green plants in the background.

Ceramic mug on a table: an example training image — Compare the example caption with the visible image.

Review the dataset in Curaset

Upload the images and their TXT files together. View each image alongside its caption. Edit a caption individually, or select several images to add, remove or replace a tag. Check how many captions will change before applying a bulk edit. If the result is wrong, undo the edit.

Check before exporting

Use the “Empty captions” filter to find missing text, then run “Check” to review the dataset. Inspect each flagged image. An unusual tag or aspect ratio may be intentional, so review it before making changes. Make sure the captions still match any images you have cropped or edited. Wait for “All changes saved” before downloading the ZIP.

Inspect the exported ZIP

Extract the archive and check the image-and-TXT pairs. If you selected a different image format for export, check the resulting file extensions. Keep your original files separately. Configure training repeats, resolution buckets and other model settings in your training tool. Curaset prepares the dataset; the quality of the trained model also depends on the images and training process.