← sources

Captions for masked training images?

source: reddit (reddit:StableDiffusion)published 2026-10-28 20:48ingested 2026-10-29 02:39

retained text

If using a masked dataset (as has now been added in Fizgig,) should the captions still include every detail of the background if it is 90% or 100% ignored? Wouldn't this confuse the training?

Bonus question: When training a character LoRA, are you keeping the entire body and clothing, or only keeping the face and hair? I don't need the model to learn the specific clothing, but probably the general body shape.

document fetched 2026-10-29 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xi03z1/captions_for_masked_training_images