Home / LoRA dataset preparation

LoRA dataset preparation

Drop training images: crop, sort out duplicates, generate AI captions and download everything as a ZIP of image and .txt files.

The AI runs on your device. The first time, your browser downloads the AI model; after that it is faster. On older phones the processing can take a while.
Drop training images here

or tap to choose. 10 to 100 images.

Choose files

Step by step

  1. First decide whether duplicate and near-identical images should be skipped. The check runs when you drop the images, not when you prepare the dataset.
  2. Drop your training images. Each one appears in the list with a preview and its pixel size; use ✕ to take single images out again.
  3. Choose size, crop, image format and the detail level of the AI captions. Add a trigger word if you need one.
  4. Tap “Prepare dataset”. The images are cropped and captioned one after another; progress is shown below the list.
  5. Read each caption in the text box next to its image and fix anything that is wrong or missing.
  6. Download everything with “Download all as ZIP”. The files are numbered and every image has a .txt file with the same name.

What you can use it for

  • Person or character: For a LoRA of yourself or an invented character, collect photos with different poses, angles and lighting. The trigger word is then placed in front of every caption.
  • Your own art style: For a style LoRA, use works made in a similar way but with different subjects. The captions should mainly describe what is shown, so the model learns the style as the common element.
  • Product or object: A specific object should later appear in new scenes. Photos from several sides in front of different backgrounds help training separate the object from its surroundings.
  • Keep mixed formats: If you have portrait and landscape images, choose “Keep aspect ratio (bucketing)”. Nothing important gets cut off at the edges.
  • Just crop and number: If you prefer writing captions yourself, set the AI to “None (write yourself)” and use only cropping, the duplicate check and the ready-made file structure.

What happens?

  • Crop: square from the center, or keeping the aspect ratio in steps of 64 (bucketing).
  • Duplicates: duplicate and near-identical images are detected on drop and skipped.
  • Captions: Florence-2 describes every image in English. You can edit each caption afterwards.
  • Trigger word: is put in front of every caption.

The ZIP contains a .txt file with the same name for each image (001.png + 001.txt) – the format Kohya_ss, OneTrainer, AI-Toolkit and SimpleTuner expect.

Frequently asked questions

Are my images uploaded?

No. Cropping and captions happen in your browser.

How many images do I need?

For people and styles 15 to 40 good, varied images are often enough. Quality matters more than quantity.

What does “smaller than target size” mean?

The image's shorter side is smaller than the selected size. It may then be enlarged during cropping and look soft. Choose a smaller target size, a better photo, or enlarge it first with “Upscale image”.

Square or bucketing, which is better?

“Square from the center” works when the subject sits in the middle. Bucketing keeps the aspect ratio and rounds both sides to multiples of 64; the long side roughly matches the selected size.

Is the trigger word added twice?

No. If a caption already starts with the trigger word, it isn't added again. Otherwise it is placed in front of the text, followed by a comma.

What happens if I change a setting after preparing?

Tap “Prepare dataset” again so cropping and format are recalculated. Existing captions are kept.

Can I enter captions myself beforehand?

Yes. If the box already contains text, no AI caption is generated for that image. That lets you describe selected images yourself.

Are my original file names kept?

No. Images are saved in list order as 001, 002 and so on. This avoids trouble with special characters in training programs.

How does the tool detect duplicates?

It compares a small fingerprint of each image. The same photo at a different size or nearly identical shots are caught this way. The status line tells you how many were skipped.