Prepare your photo dataset

A LoRA is only as good as the photos behind it. A short, well-chosen set beats a large, messy one every time.

How many photos

10 to 15 images of the same subject is what tendre.red's training screen asks for. More photos do not help if they all look alike, variety matters more than volume. Real photos and AI-generated ones can be mixed in the same set.

Vary angles and expressions

Mix close-ups, half-body shots and a few different angles (front, three-quarter, profile). Add a couple of different expressions and outfits so the LoRA learns the face, not one specific photo.

Light and framing

Pick well-lit photos where the face is sharp and clearly visible, ideally not backlit or heavily shadowed. Avoid heavy filters or big makeup changes between shots, they teach the model the wrong thing.

What to leave out

Skip blurry or duplicate shots, photos where the subject is tiny in the frame, and anything showing other people. One clean face per photo keeps training focused on the right subject.

Short on photos

Training still runs with fewer than 10 photos, even a single one, but likeness usually suffers. Add as many clean, varied photos as you can before starting, mixing in AI-generated ones of the same subject if real photos are scarce.

Ready to train

Import the set in the training screen, tendre.red auto-captions each photo (WD14 tagging) before training starts. See Train a LoRA for the full walkthrough, and Character consistency once your LoRA is trained.