The Face of Persuasion: Analyzing Bias and Generating Culture-Aware Ads

TL;DR

The work investigates demographic bias in real and text-to-image generated advertisements and examines how gender and race influence perceived persuasiveness by LLMs and multimodal models. It introduces CulGen, a culture-aware image generation framework, and demonstrates its potential to tailor ads to country-specific cultural cues while preserving the underlying ad message. The study reveals persistent racial and gender biases in both real and generated ads, and highlights the challenges of cross-cultural targeting, bias in judgments, and the need for careful, ethically guided personalization. Overall, the paper contributes a methodology for evaluating diversity and persuasion across demographics and cultures, and provides a practical path toward more culturally aware advertising content generation.

Abstract

Text-to-image models are appealing for customizing visual advertisements and targeting specific populations. We investigate this potential by examining the demographic bias within ads for different ad topics, and the disparate level of persuasiveness (judged by models) of ads that are identical except for gender/race of the people portrayed. We also experiment with a technique to target ads for specific countries. The code is available at https://github.com/aysanaghazadeh/FaceOfPersuasion

Figures (7)

• Figure 1: Selection of the more persuasive image by InternVL InternVL. Image 1 features a Black woman; Image 2 is a White woman. InternVL selected Image 2 as more persuasive. Red marks reasoning bias.
• Figure 2: CulGen for creating country-targeted ads using cultural symbols from existing ads. CA is cross-attention. The denoising condition is computed based on the time-step at the bottom of the figure, Steps 1, 2, and 3, while embeddings for Condition Scheduler are generated in upper side. MM-DiT block and noise scheduler are SD3 SD3 modules.
• Figure 3: Example on different reasoning for choosing more persuasive images.
• Figure 4: Distribution of advertisement images in PittAd dataset over different countries.
• Figure 5: Examples of cultural image generation. Action-reason prompts: (a) I should drink this beer because it is as light as feather. (b) I should use this deodorant because it is as fresh as mint.