How to Pick Images for a Character LoRA (free checklist PDF)

Most "my LoRA doesn't look like my character" problems come from the dataset, not the settings. Here's the short version of what we do for every character we train:

1. 20–25 images is the sweet spot. Fewer, better images beat many similar ones.

2. The mix: about 40% close-ups, 35% half body, 25% full body.

3. Vary everything except the character: angle, expression, light, background, pose.

4. Write down the 3–6 must-keep details and check every image. Small marks like scars, moles and notches drift first.

5. Cut blurry images, tiny faces (under 200 px), text and watermarks, other characters and near-duplicates.

6. Caption what changes, not what stays. Leave fixed features out so the trigger word learns them.

7. Test your trigger. On Z-Image, one of our l33t-style triggers sometimes printed as text on clothes, so we now test plain full names too.

The full illustrated guide with a printable checklist is attached as a PDF.

Training a character for a game, comic, brand or channel? Bravikron Labs does custom LoRAs: bravikron.com