skip to content
Site header image Nerdy Momo Cat

Your Background Remover May Just Be Setting the Background to Be Transparent

Thoughts Tips

I came across this post on reddit today where they use birefnet for background removal and then Florence for object detection, and the Florence model ends up restoring their background. No, this isn’t a training data issue. TIL, Apparently some background removal apps just set the transperancy of background pixels to 0 — they still have all the original color information in them. And all someone would need to do to restore the image would be to set all pixels to be opaque.

A pretty huge privacy risk, especially for someone like me, who uses this as an alt account

A screenshot of a reddit post asking why a model can read background from their “background removed” image, showing florence input on left with no background and complete original image on right with all information.
A screenshot of a reddit post asking why a model can read background from their “background removed” image, showing florence input on left with no background and complete original image on right with all information.