Publications
2023
Shadows Don’t Lie and Lines Can’t Bend!
Generative Models don’t know Projective Geometry…for now
Ayush Sarkar*, Hanlin Mai*, Amitabh Mahapatra*, Svetlana Lazebnik, David A. Forsyth, and Anand Bhattad
CVPR 2024
* equal contribution
Using only derived geometric properties, we can distinguish generated images from real ones without using image pixels. Thus, we show that current generative models cannot reliably generate correct shadows and/or perspective geometry.