Spot repeated frames before captioning a character dataset

Open your candidate training images as a contact sheet before writing captions. A folder with 40 files can still contain the same pose 30 times.

Group the near-duplicates and keep the clearest examples. Then check what is missing: a profile, a wider view, different lighting, or a background that is not in every image. Keep the character consistent while varying those things.

If every photo has the same coat, sofa and camera angle, more captions alone will not add the missing variation. Mark the gaps first, then add images for those gaps. Saves labelling same pose over and over.

1 comment

Spot repeated frames before captioning a character dataset