From the timeless "Mona Lisa" to the captivating "Girl with a Pearl Earring," certain artworks embed themselves in our consciousness while others fade away. While artists might cite traditional principles of memorable art, researchers have now pioneered a revolutionary approach: leveraging artificial intelligence to generate and analyze memorable imagery.
A groundbreaking study utilizing advanced machine learning algorithms has produced striking visual examples—from unforgettable cheeseburgers to easily forgotten coffee cups—that precisely illustrate what makes certain images stand out in our memory. The research revealed that images most effectively retained by human viewers shared common characteristics: vibrant color palettes, uncluttered backgrounds, and prominently centered subjects. These findings were recently unveiled at the prestigious International Conference on Computer Vision.
"Visual communication transcends linguistic boundaries," explains Phillip Isola, the Bonnie and Marty (1964) Tenenbaum CD Assistant Professor of Electrical Engineering and Computer Science at MIT and co-senior author of the study. "While extensive literature exists on memorability, our AI-driven methodology allows us to actually visualize these abstract concepts, providing concrete visual definitions for qualities previously difficult to articulate."
This innovative research builds upon an earlier framework known as MemNet, which evaluates image memorability and identifies the specific visual elements contributing to its memorability score. MemNet's predictive capabilities stem from comprehensive testing involving 60,000 images shown to human participants, who ranked them based on memorability.
The current study introduces GANalyze, a sophisticated model employing generative adversarial networks (GANs) to demonstrate the visual transformation of an image from ordinary to unforgettable. GANalyze enables viewers to witness the gradual metamorphosis of images—such as a panda obscured by bamboo foliage evolving into a commanding focal point, with its contrasting black features creating a striking impression against its white fur.
The image-modification GAN operates through three interconnected components. An evaluator module, built upon MemNet's architecture, adjusts the memorability parameters of a target image and determines the optimal modifications. A transformer module implements these adjustments, while a generator produces the final enhanced image.
The transformation process resembles a captivating time-lapse visualization. When a cheeseburger is adjusted for maximum memorability, it appears more substantial, visually appealing, and—as the researchers humorously note—more appetizing than its original version. Similarly, a ladybug gains enhanced sheen and visual intentionality. Interestingly, when optimizing a pepper's memorability, the AI unexpectedly transformed it from green to red, mimicking natural ripening processes.
The investigation also identified the most influential factors affecting image memorability. Through controlled online experiments, participants viewed images with varying memorability scores and identified duplicates they recognized. The most memorable images consistently featured closer subject perspectives, creating the illusion of larger objects or animals within the frame. Additional significant factors included enhanced brightness, centralized subject positioning, and square or circular compositional elements.
"These visual characteristics align with evolutionary adaptations in human cognition, and our AI system has successfully identified these patterns," notes study co-author Lore Goetschalckx, a visiting graduate student from Belgium's Katholieke Universiteit Leuven.
The research team also adapted GANalyze to generate images with varying aesthetic and emotional appeal. They discovered that images rated highly for aesthetic and emotional qualities shared characteristics with memorable images: increased brightness, richer coloration, and shallow depth of field creating background blur. However, the team observed that aesthetic excellence and memorability don't always correlate perfectly.
The researchers envision numerous practical applications for GANalyze technology. The system could potentially detect and even address memory impairments by enhancing object visibility within augmented reality environments.
"Rather than pharmaceutical interventions for memory enhancement, we could modify the visual environment through augmented reality to make commonly misplaced items—like keys or glasses—visually distinctive," suggests co-senior author Aude Oliva, a principal research scientist at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) and executive director of the MIT Quest for Intelligence.
GANalyze could also revolutionize educational materials by creating indelible visual aids that enhance information retention. "This technology could transform educational methodologies," Oliva asserts. Additionally, GANs are already being employed to generate synthetic, realistic images for training automated systems to recognize objects and environments they might rarely encounter in real-world scenarios.
Generative models represent an emerging frontier in human-machine creative collaboration. Alex Andonian, a graduate student at MIT's Department of Electrical Engineering and Computer Science and study co-author, has dedicated his doctoral research to exploring these technologies precisely because of their collaborative potential.
"Traditional design software allows adjustments to brightness or contrast, but not to fundamental qualities like memorability or aesthetic appeal—GANs enable this level of control," Andonian explains. "We're only beginning to uncover the full spectrum of possibilities these models offer."
The research received funding from the U.S. National Science Foundation.
