How to caption images for LoRA training
Choose a caption format that suits your model and training tool. Then compare each caption with its image. In Curaset, you can edit existing TXT files or write missing captions manually. The editor does not generate AI captions.
Choose tags or sentences
Comma-separated tags and natural-language descriptions are different caption formats. Check the documentation for your model and training setup before choosing one. A tag counter alone cannot tell you whether a caption is useful, so there is no reason to split a good sentence into tags just to raise the tag count.
Tags: ceramic mug, wooden table, soft daylight
Sentence: A ceramic mug sits on a wooden table in soft daylight.Describe details that matter
Review the subject, pose or arrangement, background and lighting. If you start with an automatically generated caption, remove any details it has invented. Capture useful differences between images instead of copying the same description across the whole folder. The right caption length depends on your model and task.
Keep your trigger word consistent
If your training setup uses a trigger word, keep its spelling consistent. Use “Dataset tags” to review tag spellings, then check that the trigger appears where you need it. Whether it should replace the subject description or appear alongside it depends on your training setup. Decide how you will use it before applying a bulk edit.
Review a bulk edit before applying it
If any images are already selected, click “Clear selection”. Filter the dataset. Click “Select matches” or select the images you want to edit individually. Choose “Add tag”, “Remove tag” or “Replace tag”, then check the selection and how many captions will change before applying the edit. These operations work with tags separated by commas or line breaks. For a caption written as a sentence, edit the text directly if a tag operation would change its meaning.
Before: ceramic mug, wood table, daylight
Replace tag: wood table → wooden table
After: ceramic mug, wooden table, daylightReview, save and export
Check for empty captions and repeated tags, then review a few images from each group you edited. Undo any incorrect bulk edit before continuing. Wait for “All changes saved”, download the ZIP and inspect the TXT files inside. Keep the original dataset so you can compare the captions before and after editing.