Dataset Preparation for LoRA Training

A good dataset is consistent, clean, and legally usable. Prefer original or licensed images, remove duplicates, and include useful variation in pose, camera, lighting, and background.

Separate concepts when possible. A dataset that mixes outfit, face, location, and style without clear captions can make the LoRA hard to control.