Week 8 Lab: Inpainting seams lab

A lab spent generating and then deliberately hunting for the seams in inpainted images, cataloguing which tells actually generalise

22 April 2027

Before the lab#

Bring three or four source images you’re comfortable editing and sharing with the group — photos of places or objects work better than photos of people for this exercise, and are simpler to get consent-free.

In the lab#

Mask a region of each image — a hand, a background object, a section of sky — and generate a fill. Then swap images with another pair and try to find the mask boundary without being told where it was. Some fills are genuinely undetectable at a glance; the interesting data is in which ones are, and which categories of region resist detection versus which give themselves away every time.

Second pass: zoom in. A flaw invisible at thumbnail size is often obvious at full resolution, which is itself worth noting — most of these images are viewed at thumbnail size in the wild.

Afterwards#

Your category list — the recurring types of tell, not the individual images — is the useful artefact here. Individual examples date quickly as the tools improve; a mechanism (“boundaries where a repeating pattern meets a non-repeating one”) tends to stay useful longer.