Cultural Representation in AI-Generated Visual Storytelling: Effects of Culturally Grounded Narrative Prompting

Recent advances in text-to-video (T2V) generation have enabled the creation of increasingly realistic and coherent visual sequences from natural language prompts [1]. As these systems are increasingly used for visual storytelling, the representation of culturally specific content has become an important concern. Previous research has shown that AI-generated images and videos may reproduce cultural elements inaccurately or omit culturally expected details [2–4]. However, less attention has been given to how the formulation of a narrative prompt may affect cultural representation across a sequence of scenes. This study addresses this question by comparing two prompting approaches: a generic narrative prompt and a culturally grounded narrative prompt. Both are used to generate the same five-scene story centred on the Onam festival in Kerala, India, while keeping the narrative structure, visual style, and generation workflow consistent. The research question being, How does culturally grounded narrative prompting affect the cultural representation of AI-generated visual stories? By comparing the same narrative under two prompting conditions, the study examines whether cultural grounding in the prompt can influence how cultural elements are represented and integrated within a visual story.

2.1 Cultural Representation and Prompt-Based Generation Prompt formulation has been identified as one factor that can shape cultural representation in generative AI. Ventura et al. [3] investigated cultural knowledge in text-to-image models through culturally oriented prompt templates structured around cultural dimensions, domains, and concepts. Their findings indicate that variations in prompt formulation can influence the cultural content produced by the models. The study, however, evaluates individual images and does not address how cultural information is maintained or expressed across a sequence of narrative scenes. The limitations of visual presence as an indicator of cultural representation are also evident in CulturalFrames [5], which reports that generated images may not satisfy either explicit or implicit cultural expectations. These findings provide a basis for examining cultural grounding at the narrative-prompt level. In the present study, culturally grounded information is incorporated directly into a narrative prompt and compared with a generic narrative prompt to examine its effect on cultural fidelity, cultural contextuality, and the integration of cultural elements within the story. 2.2 Cultural Evaluation of AI-Generated Video CultureVidBench [2] extends the evaluation of cultural representation to text-to-video (T2V) generation. Its benchmark contains 1,000 culturally grounded prompts covering 12 countries and 14 cultural aspects, including material culture, social practices and performances, and rituals and ceremonies. The results show that T2V models can achieve strong semantic adherence and visual quality while still failing to capture fine-grained cultural details, particularly in social practices and rituals [2]. While CultureVidBench evaluates cultural understanding across a broad set of prompts, it does not examine whether different approaches to embedding cultural information in a narrative prompt affect the representation of the same story. The present study addresses this gap by comparing generic and culturally grounded narrative prompts under consistent generation conditions. 2.3 AI-Assisted Visual Storytelling Generative AI has also been used to support narrative development and visual storytelling. StoryDiffusion [6] combines text-to-text and text-to-image models to turn textual narratives into visual storyboards. A user study with 12 UX design students showed how users worked between narrative development and image generation, using both user-directed and AI-directed approaches [6]. However, this work focuses mainly on storyboarding, narrative creation, and human-AI creative interaction, rather than cultural representation. The present study builds on this area by looking at how culturally grounded information in a narrative prompt can shape the cultural representation of the resulting visual story. This study connects these two research areas by examining whether explicitly embedding culturally grounded information in a narrative prompt affects the cultural representation of the resulting visual story.

3.1 Experimental Design This study uses a comparative experimental design to compare two narrative prompting conditions: generic narrative prompting and culturally grounded narrative prompting. Both conditions were used to generate the same five-scene story, “An Onam Morning” based on the Onam festival in Kerala, India. The prompting condition was the main independent variable, while cultural fidelity, cultural contextuality, and narrative - cultural coherence were used as the dependent evaluation dimensions.The story structure, main character, scene sequence, visual style, video generation model, scene duration, and motion prompts were kept consistent across the two conditions. The key difference was the cultural information included in the image prompts used to establish each scene. 3.2 Story and Scene Design A five-scene narrative, “An Onam Morning,” was created to represent a sequence of everyday activities associated with Onam and are designed to include different forms of cultural representation. The scenes follow the same character from morning preparation to a family celebration: 1. Morning preparation: The girl wakes up and begins her morning routine. 2. Temple visit: She visits a temple and prays. 3. Pookalam: She and her family arrange fresh flowers to create a pookalam. 4. Onam sadya: The family shares a traditional sadya served on banana leaves, with the characters eating with their hands. 5. Celebration: The family gathers around the completed pookalam. 3.3 Prompt Conditions For each scene, two image prompts were prepared. The generic prompt described the scene and narrative action while avoiding explicit cultural references. The culturally grounded prompt described the same narrative action while adding culturally specific information relevant to the Onam setting. For example,in Scene 4, the generic prompt described a family sharing a homemade vegetarian meal using ordinary plates and serving dishes. The culturally grounded prompt specified an Onam sadya served on banana leaves, Kerala clothing and home surroundings, traditional festive food, and eating with hands. Thus, the narrative event remained the same while the cultural information provided to the image-generation system differed.This distinction allowed us to compare whether explicit cultural grounding in the image prompt affected the resulting visual representation. 3.4 Image and Video Generation For each prompting condition, keyframe images were first generated for all five scenes using Nano Banana. The generated keyframes were then used as visual inputs for video generation with Kling AI. Each scene was generated as a five-second video using image-to-video generation. The same motion prompt was used for the corresponding generic and culturally grounded versions of each scene. The five generated clips from each condition were then arranged in the original narrative order to produce two approximately 25-second videos: one generic-prompt version and one culturally grounded version. The final comparison therefore consisted of two videos with the same five-scene narrative and corresponding scene durations, while differing in the image prompts used to establish their cultural content.

4.1 Evaluation Dimensions To compare the two prompting conditions, the generated videos were assessed at the scene level using three dimensions focused on the accuracy, contextual placement, and narrative integration of cultural elements. These dimensions were applied consistently to both the generic and culturally grounded conditions. • Cultural Fidelity refers to the accuracy of culturally specific objects, clothing, practices, and activities. • Cultural Contextuality refers to the appropriateness of cultural elements within their social and physical setting. • Narrative-Cultural Coherence refers to the integration of cultural elements with the events and progression of the narrative. 4.2 Scoring Procedure Each dimension was scored from 0 to 2, where 0 indicates absent/incorrect or disconnected representation, 1 indicates partial or weak representation, and 2 indicates clear, appropriate, or integrated representation.

This study examines whether incorporating culturally specific information into narrative prompts leads to improved cultural representation in AI-generated visual stories. The research question and hypotheses focus on three dimensions: cultural fidelity, cultural contextuality, and narrative-cultural coherence. Research Question: How does culturally grounded narrative prompting affect the cultural representation of AI-generated visual stories? To address this question, the study compares the two prompting conditions across three dimensions of cultural representation: cultural fidelity, cultural contextuality, and narrative-cultural coherence and hence the study proposes three hypotheses: • Cultural Fidelity(H1): Culturally grounded narrative prompting will result in higher cultural fidelity than generic narrative prompting. • Cultural Contextuality(H2): Culturally grounded narrative prompting will result in higher cultural contextuality than generic narrative prompting. • Narrative-Cultural Coherence(H3): Culturally grounded narrative prompting will result in higher narrative-cultural coherence than generic narrative prompting.

The comparison showed a clear difference between the two prompting conditions across all three evaluation dimensions. The culturally grounded condition received higher ratings for cultural fidelity (2 vs. 1), cultural contextuality (2 vs. 0), and narrative-cultural coherence (2 vs. 0) compared with the generic condition. The generic video provided some culturally relevant visual elements but lacked sufficient cultural context and narrative integration. In contrast, the grounded video more clearly represented culturally specific elements and integrated them with the events of the Onam story. These findings provide preliminary support for H1, H2, and H3.