What Drives Compositional Generalization in Generative Models?

Welcome to the companion website for our research paper on compositional generalization in generative models.

Results on Shapes2D

Results on Shapes2D

Results on CelebA

Results on CelebA

Results on CLEVRER-Kubric

We show results below on level-2 compositions (large red cube). The training data has the base concepts encoded separately, with different shapes taking the red color (leaving cubes to be only green) and different sizes.

Training Data

CLEVRER Setting D CLEVRER Setting E CLEVRER Setting F CLEVRER Setting G

Results with MaskGIT (Standard Training Objective)

MaskGIT CLEVRER D MaskGIT CLEVRER E MaskGIT CLEVRER F MaskGIT CLEVRER G

Results with DiT

DiT CLEVRER D DiT CLEVRER E DiT CLEVRER F DiT CLEVRER G

Results with MaskGIT (JEPA-based Training Objective)

MaskGIT JEPA CLEVRER D MaskGIT JEPA CLEVRER E MaskGIT JEPA CLEVRER F MaskGIT JEPA CLEVRER G