Episode Details

Tapping an object in Google Photos and watching it vanish feels like magic, but it's actually a two-stage machine learning pipeline. This episode breaks down the segmentation phase — where a vision transformer identifies exactly which pixels belong to the object — and the inpainting phase, where a generative model fills the hole with brand-new pixels. We explore why artifacts appear, how models like Meta's SAM and Samsung's LaMa push the limits, and why this same technology is used for privacy redaction and medical imaging.