All Roads Lead to Rome? Exploring Representational Similarities Between Latent Spaces of Generative Image Models
This work addresses the problem of understanding representation learning in generative models for researchers, but it is incremental as it builds on existing methods to explore similarities without introducing new paradigms.
The study investigated whether different generative image models learn similar underlying representations by measuring latent space similarity across VAEs, GANs, Normalizing Flows, and Diffusion Models, finding that linear maps between latent spaces preserve most visual information and that gender is the most similarly represented attribute in CelebA models.
Do different generative image models secretly learn similar underlying representations? We investigate this by measuring the latent space similarity of four different models: VAEs, GANs, Normalizing Flows (NFs), and Diffusion Models (DMs). Our methodology involves training linear maps between frozen latent spaces to "stitch" arbitrary pairs of encoders and decoders and measuring output-based and probe-based metrics on the resulting "stitched'' models. Our main findings are that linear maps between latent spaces of performant models preserve most visual information even when latent sizes differ; for CelebA models, gender is the most similarly represented probe-able attribute. Finally we show on an NF that latent space representations converge early in training.