Dataset preparation

How to prepare images for LoRA training

A LoRA can only learn the evidence in its dataset. The goal is not to collect the most pictures. It is to show the same intended subject clearly, with enough variation that the model does not memorize one frame.

How many images?

NiceVois accepts 8–40 images in any common format — JPG, PNG, WEBP, HEIC, AVIF, BMP, TIFF, or GIF. Start around 10–15 when you have genuine variety. Do not add duplicates merely to reach a higher count.

Useful variation

• Different angles and distances.
• Different backgrounds and lighting.
• Several expressions or poses for a person or character.
• Multiple views for a product.
• Different subjects that share the same treatment for a style LoRA.

What to remove

• Near-duplicate frames from the same clip.
• Images where the intended subject is tiny or blocked.
• Watermarks, screenshots with interface elements, and heavy text.
• Unrelated people, products, characters, or conflicting styles.
• Anything you do not own or have permission to use.

Resolution and format

Use valid image files at least 512 pixels wide and tall; anything that is not already JPG or PNG is converted to PNG on upload, with iPhone rotation applied. A clean, correctly framed image is more valuable than an unnecessarily large file.