Controlling memorability of face images with generative models

Area of Science:

• Computer Vision
• Cognitive Science
• Psychology

Background:

• Faces are highly memorable due to evolutionary relevance for survival and social interaction.
• Not all facial images possess equal memorability, and factors influencing this are not fully understood.
• Existing research on face memorability is limited, leaving a gap in understanding controllable image properties.

Purpose of the Study:

• To introduce a rapid and effective method for controlling the memorability of face images.
• To explore the latent space of StyleGAN for manipulating image memorability.
• To demonstrate the method's applicability to both real and synthesized images, and across different image categories.

Main Methods:

• Identified a hyperplane in StyleGAN's latent space to distinguish high and low memorable images.
• Adjusted image memorability by shifting vectors along the hyperplane's normal.
• Analyzed the impact of facial attributes on memorability within the StyleGAN latent space.
• Developed a multi-level technique to modify memorability while preserving facial identity.

Main Results:

• Successfully controlled the memorability of face images using the proposed StyleGAN-based approach.
• Demonstrated the method's effectiveness on both real-world and AI-generated faces.
• Extended the memorability modification technique to diverse image categories beyond faces.
• Validated a multi-level approach for fine-tuning memorability scores without altering identity.

Conclusions:

• The proposed method offers a novel way to manipulate image memorability efficiently.
• The technique is versatile, applicable to various image types and both real and synthetic data.
• Further research can leverage this approach for applications in media, advertising, and cognitive studies.