PVP: An Image Dataset for Personalized Visual Persuasion with Persuasion Strategies, Viewer Characteristics, and Persuasiveness Ratings
Visual persuasion, which uses visual elements to influence cognition and behavior, has gained new relevance with recent advancements in AI. However, a key bottleneck is the lack of datasets linking persuasive images with personal information about evaluators. To address this, we release the Personalized Visual Persuasion (PVP) dataset, comprising 28,454 persuasive images across 596 messages and 9 persuasion strategies. Each image is annotated with persuasiveness scores from 2,521 human annotators, along with their demographic and psychological characteristics, including personality traits and values. We demonstrate the utility of PVP by developing a persuasive image generator and automated evaluator. Our results show that incorporating psychological characteristics improves the generation and evaluation of persuasive images, offering valuable insights for personalized visual persuasion.
Visual persuasion, which uses visual elements to influence cognition, emotions, and behaviors, plays a vital role in domains such as advertising, memes, propaganda, and political communication [1]. It has been a core part of human expression throughout history, communicating power and moral values through religious and political art. Various disciplines, including communication studies and social psychology, have studied visual persuasion extensively [2, 3, 4, 5]. Recently, researchers have explored using generative models to produce persuasive visuals [6, 7]. However, a major challenge remains: persuasive effectiveness depends heavily on the psychological characteristics of the viewer, yet existing datasets rarely provide such information.
To address this gap, we construct and release the Personalized Visual Persuasion (PVP) dataset, which consists of 28,454 images tied to 596 messages designed to change behaviors (e.g., “Do not smoke”), spanning 20 real-world topics inspired by U.S. government agencies. The dataset incorporates nine persuasion strategies grounded in theory (e.g., gain frame), and images were sourced via both DALL·E generation and Google Image Search. A unique feature of the PVP dataset is that each image is annotated with persuasiveness scores provided by 2,521 human annotators, alongside their demographic data, habits, Big Five personality traits [8], and values [9, 10]. This enables a nuanced analysis of how psychological characteristics influence perceived persuasiveness and supports the development of personalized persuasive systems. Based on the PVP dataset, we conducted two tasks. In the first task, we developed an automated evaluator that takes a message (e.g., “Do not smoke”), an image, and the target viewer’s profile, and predicts the image’s persuasiveness for the target viewer. We found that including the viewer’s psychological characteristics significantly improves prediction accuracy. The performance of different GPT models varied as to whether an input image is provided as the image itself (multimodal) or its description; GPT-4o performed best on image inputs, while GPT-4o-mini excelled with text inputs. However, a small LLaMA3 model fine-tuned on PVP outperformed both. In the second task, we built a personalized image generator that creates persuasive images conditioned on a message and the target viewer’s profile. We evaluated its performance using the above evaluator. A LLaMA3 model fine-tuned on PVP achieved the highest generation performance, followed by GPT-4o-mini and GPT-4o. We also found several limitations and room for improvement. For example, the generator sometimes showed a misalignment between the image description and the intended message, as well as a poor understanding of the viewer’s psychological characteristics, particularly values. We expect that PVP will serve as a foundation for future research on personalized visual persuasion, enabling the development of more psychologically aware image generation models and improving the effectiveness and ethical alignment of persuasive technologies. This work will be published at ACL 2025.
Junseo Kim, Jongwook Han, Dongmin Choi, Jongwook Yoon, Eun-Ju Lee, Yohan Jo
Daniel Chandler and Rod Munday. 2011. A dictionary of media and communication. Oxford University Press, USA.
Paul Messaris. 1996. Visual persuasion: The role of images in advertising. sage publications.
Lawrence L Garber and Eva M Hyatt. 2003. Color as a tool for visual persuasion. In Persuasive imagery, pages 313–336. Routledge.
Kiwon Seo. 2020. Meta-analysis on visual persuasion– does adding images to texts influence persuasion. Athens Journal of Mass Media and Communications, 6(3):177–190.
Thomas Miller. 1998. Visual persuasion: A comparison of visuals in academic texts and the popular press. English for specific purposes, 17(1):29–46.
Ana Erika Ruiz-Arellano, David Abdel Mejía-Medina, Victor H. Castillo-Topete, María Berenice Fong-Mata, Ervey L. Hernández-Torres, Paloma Rodríguez-Valenzuela, and Enrique Berra-Ruiz. 2022. Addressing the use of artificial intelligence tools in the design of visual persuasive discourses. Designs, 6(6):124–124.
Yaman Kumar, Rajat Jha, Arunim Gupta, Milan Aggarwal, Aditya Garg, Tushar Malyan, Ayush Bhardwaj, Rajiv Ratn Shah, Balaji Krishnamurthy, and Changyou Chen. 2023. Persuasion strategies in advertisements. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 57–66.
Lewis R Goldberg. 2013. An alternative “description of personality”: The big-five factor structure. In Personality and Personality Disorders, pages 34–47. Routledge.
Shalom H Schwartz. 2012. An overview of the schwartz theory of basic values. Online readings in Psychology and Culture, 2(1):11.
Jesse Graham, Jonathan Haidt, Sena Koleva, Matt Motyl, Ravi Iyer, Sean P Wojcik, and Peter H Ditto. 2013. Moral foundations theory: The pragmatic validity of moral pluralism. In Advances in experimental social psychology, volume 47, pages 55–130. Elsevier.
