Keep It Simple: Evaluating Local Search-Based Latent Space Editing

Abstract

AbstractSemantic image editing allows users to selectively change entire image attributes in a controlled manner with just a few clicks. Most approaches use a generative adversarial network (GAN) for this task to learn an appropriate latent space representation and attribute-specific transformations. Attribute entanglement has been a limiting factor for previous approaches to attribute manipulation. However, more recent approaches have made significant improvements in this regard using separate networks for attribute extraction. Iterative optimization algorithms based on backpropagation can be used to find attribute vectors with minimal entanglement, but this requires large amounts of GPU memory, can lead to training instability, and requires differentiable models. To circumvent these issues, we present a local search-based approach to latent space editing that achieves comparable performance to existing algorithms while avoiding the aforementioned drawbacks. We also introduce a new evaluation metric that is easier to interpret than previous metrics.