C ULTURE S CORE: Evaluating Cultural Faithfulness in Video Generation Models Anku Rani1 Wei Dai1 Shravan Nayak2 Pattie Maes1 Mahdi M. Kalayeh3 Paul Pu Liang1 1 Massachusetts Institute of Technology 2 Mila – Quebec AI Institute Correspondence: [email protected] Abstract
arXiv:2606.07311v1 [cs.CV] 5 Jun 2026
As video generation models like Veo 3.1 and LTX-2 advance, their ability to accurately represent diverse global cultures remains a critical yet understudied frontier. Current metrics, such as VideoScore, only measure visual quality but offer no mechanism for assessing cultural faithfulness. Consequently, a model that replaces a Namaste with a handshake receives the same score as one that generates the gesture correctly. We propose C ULTURE S CORE a compositional evaluation framework that decomposes cultural faithfulness into three granular dimensions: Identity (who is represented), Context (culturally localized background), and Behavior (normative gestures and interactions). We operationalize this framework through an evaluation suite spanning 10 countries, yielding 6,180 generated videos across three stateof-the-art models. Our evaluation reveals that no current model achieves culturally faithful video generation: the best-performing model reaches only 56.8% overall C ULTURE S CORE with Behavior the most challenging dimension, which remains below 52% across all models. Furthermore, human preference rankings align directionally with C ULTURE S CORE but are inverted relative to VideoScore; the highestscoring model on visual quality was ranked last by annotators, underscoring that cultural faithfulness is an essential criterion for equitable video generation.
1
Introduction
Video Generation models like Veo3 (Google DeepMind), Sora (OpenAI, 2024), and Wan2.2 (Wan et al., 2025) are capable of simulating the physical world with high faithfulness. They are already used to generate advertisements, social media content, and short films (Hume et al., 2025). As adoption grows globally, a fundamental question arises about how faithfully these models represent the world’s cultural diversity. A system trained predominantly on Western media will generate a handshake when prompted for a greeting, not a Namaste or a Salam. These systematic biases may marginalize the cultural norms of billions of people as video generation tools spread worldwide. Ensuring that
3
Netflix
these models can accurately depict diverse cultural realities is essential for equitable access to generative AI, with implications for education, media, policy communication, and social inclusion. Research on cultural evaluation in generative AI has so far focused almost entirely on text-to-image (T2I) models. Benchmarks like CulturalFrames (Nayak et al., 2025), CuRE (Rege et al., 2025), CUBE (Kannen et al., 2024), and CultDiff (Bayramli et al., 2025) have collectively revealed that T2I models frequently miss cultural cues and default to stereotypical depictions. However, images are static, while culture is often expressed through actions in sequence, requiring evaluation frameworks that reason over motion sequences rather than single frames. A Namaste requires folding the hands and bowing in a specific order; a Salam carries its own gesture and rhythm. Images cannot capture these distinctions, and neither can evaluation frameworks designed for them. On the video side, the landscape is sparse. VideoScore (He et al., 2024) and UnifiedReward (Wang et al., 2025) measure temporal consistency, visual quality, and factual grounding. While useful for assessing general video quality, these metrics can be actively misleading when used for cultural faithfulness. For example, VideoScore assigns equal scores for videos that replace a Namaste with a handshake. To address these limitations, we propose a compositional approach to evaluating cultural faithfulness in generated videos. Prior work in video understanding has shown that complex visual scenes can be more reliably analyzed when decomposed into semantically meaningful components like actions, scenes, and objects (Ray et al., 2018). Inspired by compositionality, we propose an evaluation framework called C ULTURE S CORE that decomposes a textual prompt into three culturally grounded facets: identity (who is represented and how), behavior (culturally normative gestures, prosody, expressivity), and context (culturally situated settings and social conventions). Unlike holistic scores, this finegrained metric exposes subtle cultural mismatches, such as an incorrect greeting gesture performed in the right setting, and identifies precisely which component failed. The result is not just a faithfulness score but an interpretable diagnostic that reveals where and how a model diverges from authentic cultural representation. We operationalize this framework by building an evaluation dataset grounded in CulturalFrames (Nayak et al., 2025), a benchmark of 981 culturally validated settings
Figure 1: C ULTURE S CORE is a new compositional evaluation framework that decomposes cultural faithfulness into three dimensions: Identity (who is represented), Context (culturally localized background), and Behavior (normative gestures and expressivity). We operationalize this framework through an evaluation suite of 2,943 culturally validated prompts spanning 10 countries and 5 socio-cultural domains, yielding 6,180 generated videos and 9,289 reasoning and QA pairs across three state-of-the-art models: Veo 3.1 Fast (Google DeepMind), LTX-2 (HaCohen et al., 2024), and Wan 2.2 (Wan et al., 2025). spanning 10 countries and 5 socio-cultural domains. Each setting is decomposed into 9289 question-answer pairs across identity, behavior, and context components (see Figure 1). These QA-pairs are used to evaluate the cultural faithfulness of state-of-the-art video generation models, producing component-level accuracy scores that aggregate into an overall C ULTURE S CORE. Our evaluation reveals a striking finding: models that score highest on perceptual metrics systematically score lowest on cultural faithfulness, and vice versa — a divergence that is statistically validated by native human evaluators. Contributions. To summarize, this work makes the following contributions to the study of cultural faithfulness in video generation models. 1. We propose C ULTURE S CORE, a compositional evaluation framework that decomposes cultural faithfulness into Identity, Behavior, and Context dimensions, enabling fine-grained diagnosis of when models diverge from authentic cultural representation. 2. We construct and publicly release a culturally grounded evaluation suite of 2,943 prompts spanning 10 countries and 5 socio-cultural domains. It yields 6,180 videos and 9,289 reasoning and QA pairs, providing a reusable benchmark for future work on cultural faithfulness in video generation. 3. We present a comprehensive analysis of three stateof-the-art video generation models across 10 countries, identifying systematic cultural failure modes, including a striking inverse correlation between current video generation metrics and cultural faithfulness, validated by native human evaluators.
2
Related Work
Video generation models. Early video generation models, such as Make-A-Video (Singer et al., 2023), built on the success of text-to-image diffusion models by extending spatial U-Net architectures with temporal attention modules. A significant architectural shift occurred in 2024 with the adoption of Diffusion Transformer (DiT) backbones (Peebles and Xie, 2023; Wang et al., 2024; Lee et al., 2024). Further, OpenAI’s Sora (OpenAI, 2024) demonstrated that scaling DiTbased video models yields strong temporal consistency and semantic coherence. This catalyzed a wave of opensource DiT-based video generation models, including CogVideoX (Yang et al., 2025b), HunyuanVideo (Kong et al., 2025), LTX-Video (HaCohen et al., 2024), and Wan (Wan et al., 2025) that have made substantial progress on text-video alignment, complex motion generation, real-time video generation, and instruction following. Despite these advances, video generation models are primarily evaluated on physical realism, temporal consistency, and general prompt adherence (He et al., 2024; Wang et al., 2025). Their capacity to faithfully represent different cultures and their nuances remains entirely unstudied. Cultural representation in generative AI. Research into cultural representation in AI systems across a range of modalities, from language models (Chiu et al., 2025) to multimodal systems (Nayak et al., 2024; Bhatia et al., 2024), has shown that frontier models struggle with non-Western cultural knowledge. Later, as text-toimage (T2I) generation matured, researchers extended this inquiry to visual content creation. Benchmarks
like CUBE (Kannen et al., 2024), CuRE (Rege et al., 2025), CulturalFrames (Nayak et al., 2025), and CultDiff (Bayramli et al., 2025) revealed that T2I models frequently miss cultural cues and default to stereotypical depictions. Image transcreation (Khanuja et al., 2024) further highlighted the difficulty of the problem for image editing, where the goal was to adapt cultural content across regions. On the evaluation side, CULTIVate ’s (Malakouti et al., 2026) AHEaD metric measures cultural faithfulness across alignment, hallucination, exaggeration, and diversity dimensions. However, cultural expression in video dynamically unfolds over time through gestures and interactions that a single frame cannot capture. To the best of our knowledge, no existing benchmark or metric addresses cultural faithfulness in video generation models. Video generation evaluation metrics. Early metrics like FVD (Unterthiner et al., 2019) measured statistical similarity between generated and real video distributions, while more comprehensive benchmarks such as VBench (Huang et al., 2024a) and VBench++ (Huang et al., 2024b) improved coverage by assessing perceptual dimensions like motion smoothness and temporal consistency. Yet none measured reliably whether generated content was semantically faithful to the prompt. MLLMs, with their ability to reason over both visual content and natural language, offer a more promising path, as shown by their success in text-to-image evaluation (Ku et al., 2024). Building on this, VideoScore (He et al., 2024) and UnifiedReward (Wang et al., 2025) trained VLMs on human preference data to produce scores across visual quality and semantic faithfulness. VFEval (Song et al., 2025) further confirmed that VLMs provide reliable feedback for evaluating video generation. Yet, none of these works evaluate for cultural faithfulness. In this work, we leverage VLMs to build a compositional, QA-based evaluation framework for evaluating cultural faithfulness, exposing failure modes that current metrics miss.
3
C ULTURE S CORE: A Decomposed Evaluation Framework
Evaluating cultural faithfulness in Video Generation models presents a unique technical challenge: while standard metrics like VideoScore (He et al., 2024) or Unified Reward (Wang et al., 2025) assess visual quality, temporal consistency, dynamic degree, etc., they fail to capture the granular accuracy of cultural norms. Formally, given a cultural prompt P representing a specific socio-cultural frame, the goal is to evaluate a video V such that the identity, behavior, and context are all internally consistent with the target culture. 1. Identity refers to who is represented and how they are depicted in the generated video. This dimension focuses on the characters’ physical appearance, attire, and demographic markers. For example, the visual representation of Japanese colleagues or Chinese students should reflect distinct cultural norms
in professional or academic dress and physical characteristics. 2. Behavior encompasses what the people are doing, specifically focusing on culturally normative gestures, prosody, and expressivity. Unlike static images, video must capture the specific ”gesture and rhythm” of actions. A primary example is the cultural variation in greetings: while a model might default to a Western handshake, a culturally faithful generation would correctly depict a Namaste in India or a Salam, which require specific sequences of motion and posture. 3. Context refers to the culturally localized background, including the physical setting, environmental details, and underlying social conventions. This dimension covers both the atmospheric decor, such as Islamic calligraphy or Persian rugs, and social arrangements, such as a family seated around a dastarkhwan (floor spread) rather than a Western-style dining table. Figure 1 illustrates these dimensions through a comparison of generated content across different cultural prompts. The technical challenges of this problem are three-fold: (1) Granularity Gap: Cultural errors are often subtle and hidden within high-quality visual outputs, making holistic scores unreliable; (2) Label Dependency: Models often over-rely on explicit geographic tokens (e.g., ”India”) rather than understanding the underlying cultural concepts; and (3) Temporal Complexity: Culturally specific motions (e.g., behavior) can be difficult to generate. 3.1
Cultural Dimensions and Data Curation
Our evaluation framework C ULTURE S CORE is grounded in CulturalFrames (Nayak et al., 2025), a benchmark of 981 culturally validated prompts spanning 10 countries and 5 socio-cultural domains, originally developed for evaluating cultural representation in text-to-image models. We are the first to adapt these prompts for video generation evaluation, decomposing each into our three culturally grounded dimensions of Identity, Behavior, and Context to construct a balanced evaluation set across 10 geographic regions and 5 socio-cultural categories (see Appendix A.1 for more details). Prompts were adapted using the Gemini 3 Flash model (Google, 2026) to ensure each prompt explicitly encodes geographic and cultural identifiers. Where a country name was absent from the original prompt, geographical adjectives were appended. For example, the prompt “Couple meeting at a German sports club gathering” is decomposed as: German Couple (Identity), German meeting (Behavior), and German sports club gathering (Context). Full prompt details are provided in Appendix section C. We also provide details on how countries are similar to one another in the Appendix section A.2.
Base Pr om pt "Chinese mother organizing family festival outing"
I n pu t Dat a
Ex t en ded Pr om pt
Decom posed Cultur al Keyw or d
Cultur al Exper t Ver ified + Decom posed Guidance w ith I, B, C
Identity (I)
"Chinese mother organizing family festival outing. [Behavior]: systematically arranging items, coordinating people by gesturing, ... [Identity]: a woman of Chinese heritage, dressed in practical attire, ... [Context]: outdoor celebration with traditional Chinese decorations such as red lanterns, ..."
Behavior (B) Cultur al Exper t Ver ified
Context (C)
Vi deo Gen er at i on M odel s Veo 3.1
Wan 2.2
LTX-2
Qu est i on Ver i f i er
Cu l t u r al QA Gen er at i on s
Cu l t u r eScor e Cal cu l at i on
Gem ini-3.1-Flash
I den t i t y (I ) Is the m other w ear ing clothing that featur es a 'M andar in collar 'or 'Pankou'?
Beh av i or (B) W hen handling the 'Hongbao', does one use tw o hands w ith a light-finger ed gr ip?
Vi deo Ver i f i er Qw en3-VL-235B (1) Identify v i su al ev i den ce (2) Assess cu l t u r al accu r acy vs. em bedded descr iptions (3) Evaluate t em por al an d physi cal coh er en ce (4) Check for Wester n-centr ic defaults or st er eot ypes
Con t ex t (C) Is the gr ound com posed of an expansive, flat expanse of light gr ay gr anite paving stones?
Standar d Accur acy
Identity
Behavior
Context
Aver age
Bi n ar y Eval u at i on
Over al l Cu l t u r eScor e
Figure 2: The C ULTURE S CORE evaluation framework. Base prompts are decomposed into Identity, Behavior, and Context dimensions and expanded with elaborated meaning. Videos generated by each model are evaluated via question-answer pairs, which are verified for accuracy by a VLM before scoring. Component-level accuracy scores across the three dimensions are averaged to produce an overall C ULTURE S CORE. 3.2
Evaluation Framework
Given these sourced cultural prompts, our approach consists of four main steps: (i) Counterfactual prompt augmentation, where systematic variations of cultural prompts are generated to probe model behavior; (ii) video generation, where augmented prompts are used to generate videos, (iii) IBC decomposition, where each identity, behavior, and context dimension are evaluated; and (iv) VLM-based scoring, where VisionLanguage Models (VLMs) perform fine-grained, aspectbased question answering to quantify faithfulness in each dimension. Please refer to Figure 2. Counterfactual prompt augmentation. To systematically probe how models respond to varying levels of cultural explicitness and to identify whether failures stem from missing explicit cues or absent implicit knowledge, we design three prompt variations for each cultural scenario. Base Prompt includes the original prompt adapted from a previous work on culture (Nayak et al., 2025) that is verified by native people from that country. Extended Prompt expands the prompt into an explicit descriptive scene, mapping nuances into Identity, Behavior, and Context by elaborating its meaning from the Oxford dictionary. The extended prompt is used in our evaluation framework. Geographical Constraint Removed Prompt removes specific country names (e.g., changing “Muslim Indian family” to “Muslim family”) to test whether models have internalized cultural concepts as open-world knowledge or rely on geographic triggers to activate cultural associations. An illustrative example of this strategy for a “Halal Feast” scenario is
provided in Figure 3. A sample of these prompts was evaluated by the native residents of the countries (Refer section 4.2). Video generation models. To ensure a comprehensive evaluation of cultural faithfulness, we select three state-of-the-art video generation models: Wan2.2, LTX2, and Veo 3.1 Fast. These models represent the current frontier in high-fidelity temporal synthesis and instruction following. For LTX-2 and Wan 2.2, we generate 5-second videos across three prompt variations: a base prompt, an extended prompt for structural guidance, and a version where the specific geographical country name is removed to test implicit cultural knowledge. For Veo 3.1 Fast, we test on a subset of 10% of the data because of the API costs and generate 4-second videos. More details on the model specifications are present in the Appendix section D. IBC decomposition. To evaluate the generated videos, we use gemini-3-flash-preview (Google, 2026) to automatically generate 9,289 culturally-grounded questions and their corresponding ground-truth answers. The model is provided with human-validated prompts from Cultural Frames. We prompt the model to generate diverse reasoning pairs based on two key principles. First, we embed visual descriptions, where precise physical or spatial details are embedded directly into the question so the evaluator does not rely on implicit cultural knowledge. For example, instead of asking if a person is wearing a traditional Kimono, the model asks: Is the person wearing a traditional Kimono, characterized by
Figure 3: The Base Prompt (left) includes an explicit geographic identifier (“Muslim Indian family enjoying a halal feast at home”), producing a culturally grounded scene. The Extended Prompt (center) augments the scene with decomposed Identity, Behavior, and Context descriptions, yielding richer cultural detail, such as Islamic décor, white kurta attire, etc. The Geographical Constraint Removed Prompt (right) strips the country identifier, retaining only “Muslim family enjoying a halal feast at home”, to probe whether models have internalized cultural concepts independently of geographic triggers. left-over-right wrapped lapels and wide, square-cut sleeves?” Second, we enforce temporal grounding which explicitly probes the progression of movement across frames, such as: Does the pouring behavior begin with the vessel held low, smoothly rise to a higher elevation, and return low without breaking the stream?” This approach ensures that the evaluation is rooted in specific, verifiable cultural markers rather than generic visual aesthetics. Details are in Appendix C.4. We also add a question verification stage that filters out irrelevant questions, leading to 8811 (94.92%) valid questions, see Appendix C.5 for details. A sample of these verified questions was then annotated by native residents of the country to validate the relevance of the questions. Refer to section 4.2 for more details. VLM-based scoring. To evaluate the generated videos against our questions, we use Qwen3-VL235BA22B-Instruct (Yang et al., 2025a), a state-of-the-art vision-language model designed for complex visual understanding and temporal reasoning. Given answers generated by the model, C ULTURE S CORE averages model performance across the three decomposed dimensions of identity, behavior, and context. Prompt details present in Appendix section C.6. Native residents of the country then annotated a sample of the dataset for validating answer accuracy. Refer to section 4.2 for more details.
4
Experiments
Our evaluation is guided by the following four research questions: RQ1: To what extent do current state-of-the-art video generation models faithfully represent cultural Identity, Behavior, and Context, and does performance on existing perceptual quality metrics, such as VideoScore, align with or diverge from C ULTURE S CORE? RQ2: Does providing decomposed, culturally explicit prompt guidance improve C ULTURE S CORE, and which dimensions benefit most?
Figure 4: Average C ULTURE S CORE (%) across Identity, Behavior, and Context dimensions for three video generation models under extended prompts. Context across all the models performs better than behavior and identity. Wan 2.2 outperforms the other two for Identity and Context. RQ3: Do models rely on explicit geographic identifiers to activate cultural knowledge, or have they internalized cultural concepts as geography-independent representations? RQ4: How do C ULTURE S CORE and VideoScore each align with the cultural preferences of native human evaluators, and which automated metric better reflects human judgment of cultural faithfulness? 4.1
Results
For RQ1, we present results across two complementary views: (i) how C ULTURE S CORE varies across the Identity, Behavior, and Context dimensions under our evaluation framework (Figure 4), and (ii) how VideoScore and C ULTURE S CORE diverge when compared across models and countries (Figure 5). Dimension-level findings. Context consistently yields the highest accuracy across all models, Wan 2.2 (Wan et al., 2025) achieves 69.3%, LTX-2 (HaCohen et al., 2024) achieves 62.1%, and Veo 3.1 Fast (Google DeepMind) 59.0%. Identity is the leading second dimension for Wan 2.2 (51.3%) and LTX-2 (45.0%), while Veo scores highest on Behavior (51.9%) with Identity as its weakest dimension (41.9%). Behavior
LTX-2
Brazil S. Africa
Canada
Poland
Wan 2.2
Veo 3.1
Brazil
S. Africa
Canada
Chile Poland 2.6
2.8
3
Japan
3.2
3.4
Chile 40
3.6
50
China Japan
Iran
Germany
60
70
2.2 by +21.5% (47.8% to 69.3%), and Veo 3.1 Fast by +18.0% (41.0% to 59.0%), likely because scene-level descriptions (settings, decorations, spatial layouts) map more directly onto learnable visual features.
China
Iran
Germany
India
India
VideoScore
CultureScore
Figure 5: The inverse relationship between the average of all dimensions of VideoScore, and C ULTURE S CORE. Across all regions, the model perceived as most “cinematic” (LTX-2) consistently performs worst in cultural faithfulness, while the most accurate model (Wan 2.2) scores lowest on general quality metrics.
remains weak for Wan 2.2 (49.8%) and LTX-2 (47.0%), indicating that culturally specific motion sequences represent a persistent failure mode that prompt enrichment alone cannot resolve. Refer to Figure 4 for more details. Inverse relationship between VideoScore vs. C UL As shown in Figure 5, a striking and consistent pattern emerges across all three models. LTX2 (HaCohen et al., 2024) secures the highest VideoScore (avg. 3.34) yet ranks second lowest in C ULTURE S CORE accuracy (avg. 51.4%), with Veo 3.1 Fast scoring lowest (avg. 50.9%). Wan 2.2 (Wan et al., 2025) achieves the highest cultural accuracy on average (avg. 56.8%) while receiving the lowest VideoScore ratings (avg. 2.72), though Veo 3.1 Fast (Google DeepMind) outperforms both models on C ULTURE S CORE in Iran (51.6%) and Poland (52.0%). This inverse relationship holds across 8 of 10 countries where Wan 2.2 leads, and for two countries Veo 3.1 Fast leads, suggesting that optimizing for perceptual quality does not translate to cultural faithfulness. TURE S CORE .
Effect of extended prompt. For RQ2, we compare C ULTURE S CORE under base prompts against extended prompts, which augment each scene with explicit Identity, Behavior, and Context descriptions derived from Oxford Dictionary definitions. This tests whether cultural failures stem from insufficient prompt specificity or from deeper gaps in model knowledge. Extended prompts yield consistent gains across all three models (Veo 3.1 Fast (Google DeepMind), LTX-2 (HaCohen et al., 2024), and Wan2.2 (Wan et al., 2025)) and all three dimensions (Behavior, Identity, and Context), but the magnitude varies substantially by dimension. For Identity, LTX-2 benefits most (+18.2%, from 26.8% to 45.0%), closely followed by Wan 2.2 (+17.9%, from 33.4% to 51.3%) and Veo 3.1 Fast (+11.9%, from 30.0% to 41.9%). The smaller gain for Veo 3.1 Fast suggests it already captures some object-level identity cues from base prompts. Refer to Figure 6 and Appendix section D Table 9 and 10. The Context dimension shows the largest absolute gains across all models. LTX-2 improves by +25.5% (36.6% to 62.1%), Wan
Behavior is the most resistant dimension to prompt enrichment. Despite receiving the same structured guidance, behavior gains are comparable across all models: Wan 2.2 (+15.8%), Veo 3.1 Fast (+16.2%), and LTX-2 (+16.7%). Even under extended prompting, no model exceeds 52% on Behavior, compared to 69.3% on Context for the best-performing model (Wan 2.2), indicating that temporally coherent motion sequences remain a persistent failure mode that prompt enrichment alone cannot resolve. Explicit vs. implicit cultural knowledge. For RQ3, we study whether models have internalized cultural concepts as open world knowledge or rely on geographic tokens as triggers. To answer this, we remove country names from base prompts while retaining cultural identifiers (e.g., “Indian women greeting people with Namaste.” to “Women greeting people with Namaste”). We filter for cases where the model correctly processed the base prompt (see Appendix C.7 for processing details), yielding 48.4% of prompts, and evaluate C UL TURE S CORE across all three dimensions. Removing geographic identifiers causes a consistent and substantial drop. For Veo 3.1 Fast, compared to extended prompts, removing geographic anchors causes accuracy to fall by 27.6pp on Context (59.0% → 31.4%), 23.0pp on Behavior (51.9% → 28.9%), and 16.5pp on Identity (41.9% → 25.4%). For LTX-2, compared to extended prompts, removing geographic anchors causes accuracy to fall by 33.5pp on Context (62.1% → 28.6%), 25.8pp on Identity (45.0% → 19.2%), and 20.3pp on Behavior (47.0% → 26.7%). This suggests that models rely heavily on explicit geographic tokens as cultural triggers rather than having internalized the underlying cultural concepts, and that cultural understanding in current video generation models remains largely surface-level. The effect is most severe for Chinese and Iranian prompts, where LTX-2 shows the largest drops across all three dimensions (China: −10.6 pp avg, Iran: −9.9 pp avg) when geographic anchors are removed, suggesting that low resource or visually distinctive cultures in training data are disproportionately affected. Notably, Canada and Chile show negligible drops (−0.2 pp and −0.8 pp, respectively), possibly due to greater overlap with Western training data defaults. Taken together, these two achievements reveal a consistent picture: cultural faithfulness in current video generation models is both prompt-sensitive and geographically dependent. Structured guidance improves context and identity considerably, but behavior remains stubbornly below the ceiling. We report the numbers in Figure 7.
Figure 6: Average C ULTURE S CORE (%) for base and extended prompt across the models. Extended prompts consistently outperform the base prompt across all dimensions implies extended prompting yields better output. are in Appendix B.
Figure 7: Average C ULTURE S CORE of LTX-2 and Veo 3.1 Fast across IBC dimensions with all three prompt variants. Across all dimensions, the extended prompt performs the best, with the Context dimension and the LTX-2 model benefiting the most. No country prompt exhibits the worst performance, with the greatest reduction in the Context dimension and with the Veo 3.1 Fast model.
4.2
Human Evaluation
To validate whether the divergence between VideoScore and CultureScore reflects genuine human perception, we conducted a systematic human evaluation study on Prolific with N=45 native residents (5 per country) spanning 9 countries (all except Iran, given the current geopolitical situation). Each participant evaluated 9 video triplets, yielding 135 rank observations per model. Participants were required to be verified current residents who also hold nationality of the evaluated country on Prolific, ensuring cultural authenticity of judgments. Participants spanned diverse professional backgrounds including entrepreneurs, musicians, homemakers, policymakers, and data analysts, to avoid skew toward technical AI expertise. For each triplet, participants completed a three-tier assessment: (i) Question relevance: Judging whether the C ULTURE S CORE-generated evaluation question was a meaningful probe for the given cultural context, (ii) Perceptual accuracy: Answering the Identity, Behavior, or Context question based strictly on visual evidence in the videos, and (iii) Preference ranking: Ranking videos generated from all three models from 1st to 3rd based on overall cultural faithfulness. Full annotation details
Inter-annotator agreement. Given the categorical nature of Question Relevance, we report Gwet’s AC1 (Gwet, 2008) for question relevance, as it provides a more reliable agreement estimate under skewed marginal distributions. For Perceptual Accuracy, where annotators selected all videos for which a condition held true, we report average pairwise Jaccard similarity (Jaccard, 1912) to account for the multi-label nature of responses. For preference ranking, we report Spearman’s ρ (Spearman, 1961) as a descriptive measure of ordering agreement between human preferences and evaluation metrics. Question relevance. Across all nine countries, context and identity questions were judged highly relevant on average (84.2% and 78.3%, respectively), while behavior questions received notably lower ratings (69.0%). Behavior relevance showed the greatest cross-cultural variability: German raters rated behavior questions most relevant (93.3%), while Japanese raters gave the lowest rating (44.4%). Indian raters showed a strong preference for context and identity (94.4% each) over behavior (66.7%), a pattern echoed in Japan (94.4% context, 72.2% identity vs. 44.4% behavior). Inter-rater agreement was moderate on average (AC1 ≈ 0.50). Questions marked irrelevant by a majority of raters were excluded from further analysis. Perceptual accuracy. Identity questions yielded the highest answer accuracy across all nine countries (avg. 63.0%), while behavior and context questions were comparably harder (53.7% and 51.2%, respectively), consistent with behavior’s role as the weakest dimension in our automated evaluation. Identity accuracy was especially high for German and Indian raters (88.9% each), whereas Canadian and Japanese raters found identity the most difficult (33.3% and 44.4%). Behavior accuracy showed the greatest cross-cultural variability, ranging from 33.3% (Brazil, South Africa) to 77.8%
Figure 8: Correlation between C ULTURE S CORE and native human preference rankings, across all nine countries. VideoScore (left) is inversely correlated with human preference; the model receiving the highest video score (LTX-2, red) is consistently ranked lowest by human evaluators. C ULTURE S CORE (right) is more directionally aligned with preference, with Wan 2.2 (blue) ranking highest on both cultural accuracy and human preference. Each point represents one model aggregated across all countries; error bars denote ±1 standard error. (Germany). Context accuracy diverged notably across cohorts: Indian and Polish raters scored highest (66.7% each), whereas Brazilian and Chinese raters scored the lowest (33.3% and 44.4%). Since annotators selected a set of videos per question, we report inter-rater agreement via average pairwise Jaccard similarity; agreement was moderate overall (J = 0.63), ranging from J = 0.47 (South Africa) to J = 0.78 (India). Preference ranking and metric correlation. For each prompt, annotators watched three AI-generated videos side-by-side and assigned a rank to each model (1 = most preferred, 3 = least preferred). We aggregate preferences by averaging per-model ranks across all raters and prompts within each country, then pooling across all nine countries (n = 135 rank observations per model). Human preference followed a clear ordering: VEO 3.1 Fast was most preferred (avg. rank 1.66 ± 0.07), followed by Wan 2.2 (1.94 ± 0.06), and LTX-2 was ranked lowest (2.39 ± 0.07). To assess whether automated metrics align with this ordering, we compare each model’s aggregate C UL TURE S CORE and VideoScore against its average human preference rank. C ULTURE S CORE here is computed as the fraction of culturally grounded evaluation questions for which the model’s generated answer matches the majority human ground truth; VideoScore is the average of five perceptual dimensions (He et al., 2024). VideoScore does not reflect human preference: LTX-2 received the highest VideoScore (3.36) yet was ranked last by annotators, while Veo 3.1 Fast and Wan 2.2 scored substantially lower (2.78 and 2.72, respectively) but were preferred by native human annotators (ρ = −0.50). This inversion confirms that perceptual quality metrics can actively mislead model selection. C UL TURE S CORE , by contrast, rewards WAN 2.2 most highly on the human-evaluation subset (0.656, ground truth determined by majority annotator vote across n = 27 prompts), consistent with its second-place hu-
man preference ranking. Together, these results indicate that general-purpose video quality metrics are not merely insufficient for measuring cultural faithfulness, but can actively mislead model selection when cultural accuracy is the priority.
5
Conclusion
We presented C ULTURE S CORE, a compositional framework that decomposes cultural faithfulness in video generation into Identity, Behavior, and Context dimensions. Across 2,943 prompts spanning 10 countries and 5 socio-cultural domains, existing quality metrics such as VideoScore actively mislead cultural evaluation: the model ranked highest on visual quality (LTX-2) ranked lowest on cultural accuracy, validated by native human evaluators. Decomposed cultural guidance yields meaningful gains, with Context most responsive to prompt enrichment and Behavior the most persistent failure mode. Identity-level questions emerge as the strongest predictor of human cultural preference, suggesting that who is represented and how is the dimension audiences notice first. We hope C ULTURE S CORE provides a reusable foundation for auditing cultural representation in generative AI.
6
Limitations
Our evaluation has three primary limitations. Asymmetric model coverage. Due to the high operational costs of the Veo 3.1 Fast API, we evaluate a stratified sample of 294 videos rather than the full 2,943-prompt suite used for Wan 2.2 and LTX-2. While the sample is designed to maintain cross-cultural and cross-category coverage, direct comparisons between Veo 3.1 and the other two models should be interpreted with this asymmetry in mind. Geographic scope of human evaluation. Our human evaluation is limited to native annotators from all the countries except Iran given the current geopolitical situation. We aim to include it in the future. VLM-as-judge biases. C ULTURE S CORE relies on Qwen3-VL to evaluate generated videos against culturally grounded questions. Although we include a question verification stage to reduce hallucinated or inaccurate evaluation criteria, the scoring model may itself carry cultural biases that systematically favor or penalize certain representations. We recommend treating C ULTURE S CORE as a complementary metric to be used alongside human evaluation, rather than as a standalone ground truth.
7
Ethics Statement
This work involves human evaluation of culturally sensitive video content. Annotators were recruited from Prolific as native members of the cultures being evaluated and participated voluntarily. No personally identifiable information was collected during the annotation process. The cultural prompts used in this study are grounded in the Cultural Atlas (Cultural Atlas, 2016), a publicly available resource developed collaboratively with community members. We acknowledge that cultural representation is inherently contested and that no benchmark can fully capture the diversity within any national or regional group. C ULTURE S CORE reflects cultural norms as documented in reference materials and validated by native evaluators; it should not be interpreted as a definitive or exhaustive account of any culture. We also acknowledge that the VLMs used in our automated evaluation pipeline, including Qwen3-VL and Gemini Flash, may themselves carry cultural biases that influence scoring. We further acknowledge that our operationalization of the Identity dimension, which is grounded in physical appearance, attire, and demographic markers, risks reinforcing stereotypes. Especially, the Identity dimension is a deeply complex concept extending well beyond visual surface features, encompassing language, lived experience, class, and history that a generated video cannot fully capture. Our choice to evaluate Identity through visually legible markers is a pragmatic constraint of the video evaluation setting, not a theoretical claim about what identity means. We recommend that C ULTURE S CORE be used alongside, rather than as a
replacement for, human judgment, particularly when evaluation findings are used to make claims about specific communities. The dataset and evaluation suite will be released publicly to enable replication and extension by the broader research community.
8
LLM Usage Disclosure
We used LLMs for minor writing assistance, including grammar correction and language polishing. The core research ideas, methodology, experimental design, implementation, analysis, and conclusions were developed and carried out by the authors. No LLM was used to generate research ideas, experimental results, figures, or evaluations.
References Zahra Bayramli, Ayhan Suleymanzade, Na Min An, Huzama Ahmad, Eunsu Kim, Junyeong Park, James Thorne, and Alice Oh. 2025. Diffusion models through a global lens: Are they culturally inclusive? Preprint, arXiv:2502.08914. Mehar Bhatia, Sahithya Ravi, Aditya Chinchure, EunJeong Hwang, and Vered Shwartz. 2024. From local concepts to universals: Evaluating the multicultural understanding of vision-language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6763–6782, Miami, Florida, USA. Association for Computational Linguistics. Yu Ying Chiu, Liwei Jiang, Bill Yuchen Lin, Chan Young Park, Shuyue Stella Li, Sahithya Ravi, Mehar Bhatia, Maria Antoniak, Yulia Tsvetkov, Vered Shwartz, and Yejin Choi. 2025. Culturalbench: A robust, diverse, and challenging cultural benchmark by human-ai culturalteaming. Preprint, arXiv:2410.02677. Cultural Atlas. 2016. About the cultural atlas. Accessed: 2026-03-31. Google. 2026. Gemini 3 flash preview. https://ai.google.dev/gemini-api/ docs/models/gemini-3-flash-preview. Large language model; accessed March 31, 2026. Google DeepMind. Veo: a text-to-video generation system. Technical report, Google DeepMind. Kilem Li Gwet. 2008. Computing inter-rater reliability and its variance in the presence of high agreement. British Journal of Mathematical and Statistical Psychology, 61(1):29–48. Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. 2024. Ltxvideo: Realtime video latent diffusion. Preprint, arXiv:2501.00103.
Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, and 1 others. 2024. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 2105–2123. Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, and 1 others. 2024a. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807–21818. Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, Ying-Cong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. 2024b. Vbench++: Comprehensive and versatile benchmark suite for video generative models. Preprint, arXiv:2411.13503. Tom Hume, Matthew Carey, and Thomas Iljic. 2025. Meet Flow: AI-powered filmmaking with Veo 3. Paul Jaccard. 1912. The distribution of the flora in the alpine zone. 1. New phytologist, 11(2):37–50. Nithish Kannen, Arif Ahmad, Marco Andreetto, Vinodkumar Prabhakaran, Utsav Prabhu, Adji Bousso Dieng, Pushpak Bhattacharyya, and Shachi Dave. 2024. Beyond aesthetics: cultural competence in text-toimage models. In Proceedings of the 38th International Conference on Neural Information Processing Systems, NIPS ’24, Red Hook, NY, USA. Curran Associates Inc. Simran Khanuja, Sathyanarayanan Ramamoorthy, Yueqi Song, and Graham Neubig. 2024. An image speaks a thousand words, but can everyone listen? on image transcreation for cultural relevance. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 10258–10279, Miami, Florida, USA. Association for Computational Linguistics. Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, and 33 others. 2025. Hunyuanvideo: A systematic framework for large video generative models. Preprint, arXiv:2412.03603. Max Ku, Dongfu Jiang, Cong Wei, Xiang Yue, and Wenhu Chen. 2024. Viescore: Towards explainable metrics for conditional image synthesis evaluation. Preprint, arXiv:2312.14867. Taegyeong Lee, Soyeong Kwon, and Taehwan Kim. 2024. Grid diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8734–8743.
Sina Malakouti, Boqing Gong, and Adriana Kovashka. 2026. Culture in action: Evaluating text-to-image models through social activities. In The Fourteenth International Conference on Learning Representations. Shravan Nayak, Mehar Bhatia, Xiaofeng Zhang, Verena Rieser, Lisa Anne Hendricks, Sjoerd Van Steenkiste, Yash Goyal, Karolina Stanczak, and Aishwarya Agrawal. 2025. CulturalFrames: Assessing cultural expectation alignment in text-to-image models and evaluation metrics. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 20918–20953, Suzhou, China. Association for Computational Linguistics. Shravan Nayak, Kanishk Jain, Rabiul Awal, Siva Reddy, Sjoerd Van Steenkiste, Lisa Anne Hendricks, Karolina Stanczak, and Aishwarya Agrawal. 2024. Benchmarking vision language models for cultural understanding. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5769–5790, Miami, Florida, USA. Association for Computational Linguistics. OpenAI. 2024. Video generation models as world simulators. William Peebles and Saining Xie. 2023. Scalable diffusion models with transformers. Preprint, arXiv:2212.09748. Jamie Ray, Heng Wang, Du Tran, Yufei Wang, Matt Feiszli, Lorenzo Torresani, and Manohar Paluri. 2018. Scenes-objects-actions: A multi-task, multi-label video dataset. In Proceedings of the European conference on computer vision (ECCV), pages 635–651. Aniket Rege, Zinnia Nie, Mahesh Ramesh, Unmesh Raskar, Zhuoran Yu, Aditya Kusupati, Yong Jae Lee, and Ramya Korlakai Vinayak. 2025. Cure: Cultural gaps in the long tail of text-to-image systems. Preprint, arXiv:2506.08071. Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics. Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. 2023. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations. Tingyu Song, Tongyan Hu, Guo Gan, and Yilun Zhao. 2025. VF-eval: Evaluating multimodal LLMs for generating feedback on AIGC videos. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 21126–21146, Vienna, Austria. Association for Computational Linguistics.
Charles Spearman. 1961. The proof and measurement of association between two things. Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. 2019. Towards accurate generative models of video: A new metric and challenges. Preprint, arXiv:1812.01717. Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, and 43 others. 2025. Wan: Open and advanced large-scale video generative models. Preprint, arXiv:2503.20314. Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. 2024. A recipe for scaling up text-to-video generation with text-free videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6572–6582. Yibin Wang, Yuhang Zang, Hao Li, Cheng Jin, and Jiaqi Wang. 2025. Unified reward model for multimodal understanding and generation. arXiv preprint arXiv:2503.05236. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388. Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Yuxuan.Zhang, Weihan Wang, Yean Cheng, Bin Xu, Xiaotao Gu, Yuxiao Dong, and Jie Tang. 2025b. Cogvideox: Text-to-video diffusion models with an expert transformer. In The Thirteenth International Conference on Learning Representations.
Appendix This section provides additional examples to assist in the understanding and interpretation of the research work presented. Section A: Dataset Statistics Section B: Annotation Section C: Prompts Section D: Experiments
A
Dataset Statistics and Analysis
Dimension
Category
Brazil
Canada
Chile
China
Germany
India
Iran
Japan
Poland
S. Africa
Identity
Dates of Sig. Etiquette Family Greetings Religion
19 11 13 11 16
23 15 14 10 12
20 12 12 10 10
23 13 14 10 12
21 17 12 9 8
28 18 11 12 17
19 19 11 12 10
22 16 11 13 15
17 13 10 11 7
23 18 0 12 16
Context
Dates of Sig. Etiquette Family Greetings Religion
28 20 15 14 17
31 20 17 11 13
32 22 17 12 11
36 28 17 12 14
33 21 13 12 11
37 24 14 13 18
30 25 14 13 11
35 27 14 15 16
34 19 15 11 14
35 26 0 14 16
Behavior
Dates of Sig. Etiquette Family Greetings Religion
26 21 15 14 16
27 20 15 11 12
29 23 17 12 11
33 27 17 13 13
32 22 12 11 10
37 23 14 13 18
28 24 13 13 10
33 25 14 13 14
32 19 14 11 11
34 26 0 14 15
Table 1: Number of unique Identity, Context, and Behavior per country and category. South Africa contains no Family entries, reflecting the absence of Family category data points in the source CulturalFrames dataset.
A.1
Cultural Atlas
The Cultural Atlas is a landmark collaborative initiative developed through a partnership between Mosaica, SBS, and Multicultural NSW. Established in 2016, the project serves as a comprehensive educational resource aimed at enhancing cross-cultural literacy. It provides nuanced insights into the attitudes, social norms, and communication styles of Australia’s diverse populations. By synthesizing qualitative cultural observations with contemporary demographic statistics and settlement trends, the Atlas equips individuals and organizations with the tools necessary to navigate a pluralistic society. Ultimately, the project seeks to strengthen social cohesion and improve outcomes for all participants in Australia’s multicultural landscape (Cultural Atlas, 2016). To ensure a holistic understanding of each culture, the Atlas categorizes information into the following key domains: Dates of Significance Identifying pivotal religious, national, and cultural observances. This section provides historical context for community celebrations and aids in understanding the timing of significant cultural milestones. Religion An exploration of the spiritual frameworks and belief systems that shape a culture’s worldview. This includes how faith influences daily life, ethics, and community structures. Etiquette Guidelines on the social protocols and “unwritten rules” of interaction. This domain covers body language, gift-giving, and social taboos to facilitate respectful engagement. Family An analysis of the foundational social unit, focusing on kinship structures, gender roles, and authority dynamics. It examines the balance between individualistic and collectivistic values. Greetings A summary of appropriate verbal and non-verbal salutations. This includes the use of honorifics, physical gestures, and the level of formality required during initial contact.
A.2
Similarity between countries on different categories
To study the similarity between countries and different categories, We obtained cultural reference texts. We developed a Python web scraper that programmatically downloads cultural descriptions from the Cultural Atlas (culturalatlas.sbs.com.au), a publicly available cultural information resource. The script systematically retrieves articles for 10 countries (Brazil, Canada, Chile, China, Germany, India, Iran, Japan, Poland, and South Africa) across five cultural categories: greetings, etiquette, family, religion, and dates of significanceyielding up to 50 country–category text documents. For each page, the script fetches the raw HTML, strips non-content elements (e.g., navigation, scripts, and styling), and extracts the main article text from paragraph, list, and heading elements using BeautifulSoup. The extracted texts are saved as individual plain-text files, with metadata including the source country, category, and URL. To quantify cultural similarity across nations, we computed semantic embeddings of the Cultural Atlas groundtruth texts using the all-MiniLM-L6-v2 (Reimers and Gurevych, 2019) sentence transformer model. Each country’s cultural descriptionsspanning greetings, etiquette, family, religion, and dates of significancewere encoded into dense vector representations and averaged to produce a single composite embedding per country. Pairwise cosine similarity was then computed across all 10 countries to construct a country-by-country similarity matrix, visualized as a heatmap. Additionally, per-category cross-country similarity matrices were generated to examine how cultural proximity between nations varies across specific domains. This embedding-based approach enables a data-driven comparison of cultural textual content, revealing which countries share the most and least similar cultural profiles according to the Cultural Atlas reference material. The embedding-based cross-country analysis revealed that cultural similarity varies substantially across domains. Greetings emerged as the most universally convergent category, with the highest average pairwise similarity (0.608), as cultures broadly describe similar conventions around handshakes, eye contact, and formal titles. In contrast, religion was the most divisive category (mean 0.445), exhibiting the widest similarity range (0.255–0.651), reflecting the fundamental diversity of religious traditions across the sampled nations. Family structures (mean 0.558) also showed relatively high cross-cultural similarity, as most cultures describe patriarchal structures, elder respect, and extended family systems, whereas etiquette (0.415) diverged sharply, suggesting that everyday social manners are more culturally specific than familial organization. At the country-pair level, India and Japan recorded the single highest cross-country similarity score in any category (0.732 in greetings), driven by shared conventions such as bowing, hierarchy-based greeting depth, honorific suffixes, and the avoidance of physical contact with strangers. Notably, large similarity swings were observed within the same country pair across categories. For instance, Brazil and Canada scored 0.646 in greetings but only 0.299 in etiquette, a swing of 0.348, indicating that while both cultures greet similarly through handshakes and informality, their broader social norms diverge considerably. Similarly, Chile and India scored 0.656 in family but only 0.322 in religion, reflecting shared extended family structures alongside completely divergent religious traditions. Several country-level patterns emerged as particularly noteworthy. Canada, despite its multicultural identity, was not a cultural middleman but rather an outlier in etiquette, recording the lowest average similarity to all other countries (0.369) in that category. Its informal, first-name-basis culture proved genuinely unique rather than a blend of other cultural norms. Chile stood out as the most culturally distinctive nation in both religion and dates of significance, likely driven by its Easter Island traditions such as the Tapati festival and Rapa Nui culture, combined with its specific Catholic-indigenous syncretism. Conversely, Germany emerged as the most central country in religion and dates of significance, with its cultural descriptions serving as a kind of baseline against which other nations are measured, likely because the Cultural Atlas frames German culture in broadly Western terms that partially overlap with many other cultures. Finally, the intra-region similarity gap was largest for religion, where pairs such as Germany–Poland, Brazil–Chile, and China–Japan shared religious traditions within their respective regions, while cross-region pairs like Chile–Iran (0.255) were almost entirely dissimilar. Please refer to Figure 9.
(a) Greetings
(b) Dates of Significance
(c) Etiquette
(d) Family
(e) Religion
Figure 9: Heatmaps showing similarity between countries across different cultural categories.
B
Annotation
B.1
Participant Recruitment
We recruited a total of N = 45 native evaluators (5 each) across 9 countries (India, China, South Africa, Japan, Brazil, Chile, Canada, Poland, and Germany). Participants were sourced through prolific, with the explicit requirement that each participant be a native resident or citizen of the country they were evaluating. Participants spanned diverse professional backgrounds, including entrepreneurs, musicians, homemakers, policymakers, and data analysts, to ensure that cultural judgments were not systematically skewed by technical familiarity with generative AI systems. Participation was voluntary and anonymous. Participants were paid 5USD for completing the study. No personally identifiable information was collected at any stage; all responses were used solely for academic research purposes. B.2
Annotation Platform and Setup
Annotations were collected through the Prolific and Qualtrics platforms. Participants were required to complete the study on a laptop with a stable internet connection to ensure reliable video playback, and were instructed to work in a quiet environment where they could focus on visual details. The estimated completion time was 15 minutes. Each survey began with an informed consent question; participants were required to confirm that they understood and agreed to the terms of participation before proceeding. A critical instruction was provided at the outset: participants were explicitly told to base all their judgments on video frames only, and not to consider audio in their assessments. This ensured that all evaluations were grounded purely in the visual cultural content generated by each model, which is the modality C ULTURE S COREmeasures. B.3
Annotation Task Design
Each participant was assigned 9 video triplets, with each triplet corresponding to a single cultural prompt. Within each triplet, the three videos were generated by LTX-2, Wan 2.2, and Veo 3.1 Fast, respectively, and presented in randomized order to prevent position bias. For each triplet, participants completed a three-tier assessment in the following fixed order, as illustrated in Figure 9: (i) Question Relevance. Participants were shown a C ULTURE S COREgenerated evaluation question drawn from the Identity, Behavior, or Context dimension alongside all three videos. They judged whether the question constituted a meaningful and appropriate probe of cultural faithfulness for the given prompt, responding with a binary Yes/No. Questions marked as irrelevant by a majority of raters within a cohort were excluded from further analysis. (ii) Answer Accuracy. Participants were then asked to answer the same cultural question based strictly on visual evidence in the videos, selecting all videos for which the answer was affirmatively Yes. Multiple selections were permitted if the condition held true for more than one video. This tier assesses whether the cultural marker in question is visually legible to a native observer, independently of overall aesthetic preference. (iii) Human Preference Ranking. Finally, participants ranked all three videos from 1st to 3rd based on overall cultural faithfulness to the given prompt, using a drag-and-drop interface. Participants were explicitly instructed to rank based on cultural accuracy rather than visual quality or cinematic appeal. Participants were required to complete all three tiers before proceeding to the next triplet. Figure 10 presents a screenshot from the annotation platform.
Figure 10: Annotation Platform-Humans annotated samples based on Question Relevance, Answer Accuracy, and Human Preference Ranking.
C
Prompt Specifications
In the following, we provide details on all the prompts used in this research.
C.1
Identity, Behavior, Context
The first stage decomposes each base cultural prompt into its three constituent dimensions: Identity, Behavior, and Context. This decomposition is a prerequisite for all downstream stages, as it provides the structured inputs used in extended prompt construction.
System Prompt: IBC ExtrBehavior Analyze the image prompt and extract Identity, Behavior, and Context using these rules: 1. Identity: The primary human or living subject(s). 2. Behavior: The physical activity, gesture, or movement. 3. Context: The setting, atmosphere, and background details (including what is being watched). EXAMPLES: Prompt: ”Indian family greeting elder with feet-touching gesture at home” Output: {"Identity": ”Indian family”, "Behavior": ”Greeting elder with feet-touching gesture”, "Context": ”Indian home”} Prompt: ”Germans watching New Year’s concert on television” Output: {"Identity": ”Germans”, "Behavior": ”Watching”, "Context": ”New Year’s concert on television”} TASK: Prompt: ”{prompt text}” Return ONLY a JSON Identity with keys "Identity", "Behavior", and "Context". ”””
C.2
Extended Prompting
This stage constructs the Extended Prompt variant used as the primary evaluation condition in our experiments. Extended prompts augment the base prompt with explicit, dictionary-grounded visual descriptions for each of the three IBC dimensions, giving video generation models detailed guidance on what Identity, Behavior, and Context should look like in the generated video.
System Prompt: Extended Prompting task prompt = f””” TASK: Create a detailed video generation prompt. You must extend the ’Behavior’, ’Identity’, and ’Context’ using their formal definitions grounded in the Oxford Dictionary. INPUTS: • Prompt: {prompt text} • Behavior: {Behavior} • Identity: {Identity} • Context: {Context} FORMAT: ”Cinematic shot. {prompt text}. [Behavior]: [Detailed visual description of the Behavior: {Behavior} from oxford dictionary], [Identity]: [Detailed visual description of the Identity: {Identity} from oxford dictionary], and [Context]: [Detailed visual description of the Context: {Context} from Oxford Dictionary].” OUTPUT: Return ONLY the final string following the format above. Do not include labels like ’Output:’ or extra commentary. ”””
C.3
Geographical Constraint Removed Prompting
This stage constructs the Geographical Constraint Removed variant, used exclusively in the implicit cultural knowledge ablation (RQ3). This prompt variant is produced by stripping explicit country-level identifiers from the base prompt while preserving all cultural, religious, and lifestyle descriptors that do not name a specific nation. System Prompt: Geographical Constraint Removed Prompting task prompt = """ TASK: Rewrite the following image prompt to remove any mention of specific countries (e.g., India, Indian, Germany, German, USA, American, etc.). IMPORTANT: Keep religious, cultural, or lifestyle descriptors (e.g., Muslim, Halal, Festive, Buddhist, Vegan). EXAMPLES: Input: “Indian Muslim family enjoying a halal feast at home” Output: “Muslim family enjoying a halal feast at home” Input: “German elderly couple walking in a park” Output: “Elderly couple walking in a park” Input: “A traditional Japanese tea ceremony in Tokyo” Output: “A traditional tea ceremony in Tokyo” TASK: Input: “{prompt text}” Return ONLY the rewritten string. No explanations. """
C.4
Question Generator
This stage generates the culturally grounded question-answer pairs used to evaluate generated videos. For each extended prompt, a language model acting as a cultural anthropologist and visual evaluator produces a set of Yes/No evaluation questions decomposed across the Identity, Behavior, and Context dimensions. In total, this stage produces 9,289 question-answer pairs across all prompts and models.
System Prompt SYSTEM PROMPT = ””” You are an expert cultural anthropologist, visual evaluator, and prompt engineer. Your task is to analyze a given video generation prompt and systematically generate an exhaustive evaluation framework to measure the cultural faithfulness of the resulting generated video. Decompose the prompt into three categories: Context, Identity, and Behavior. Strict Guidelines: 1. Yes/No Format: Every question must be structured so that a ”Yes” indicates cultural faithfulness and a ”No” indicates a failure or cultural inaccuracy. 2. Embedded Visual Descriptions: Embed precise visual, physical, or spatial descriptions directly into the question. Do not assume the evaluating model has implicit cultural knowledge. (e.g., Instead of ”Is the person wearing a traditional Kimono?”, ask ”Is the person wearing a traditional Kimono, characterized by left-over-right wrapped lapels and wide, square-cut sleeves?”). 3. Temporal Grounding (Crucial for Behaviors): Questions MUST explicitly probe the sequence, duration, physics, or progression of the movement across frames. (e.g., ”Does the pouring Behavior begin with the vessel held low, smoothly rise to a higher elevation, and return low without breaking the stream?”). 4. Explicit vs. Implicit: Generate questions for explicitly requested elements, AND mandatory implicit elements required for cultural authenticity. Actively avoid Western-centric stereotypes (e.g., implicitly checking that a traditional daily Context avoids hyper-exoticized or religious backdrops unless the Context or the prompt requires them). 5. Weighting Strategy (1-10 Scale): • 8–10: Critical explicit elements or absolute cultural boundaries. • 4–7: Expected contextual elements and secondary Identitys that enhance authenticity. • 1–3: Minor background details or high-fidelity nuances. ”””
C.5
Question Verifier
This stage filters out questions that, despite being generated from culturally validated prompts, make claims about cultural elements that are factually inaccurate or not genuinely grounded in the target culture. This step is necessary because the question generator in Stage can occasionally produce questions that embed incorrect or hallucinated cultural details. For example, attributing a practice to a country where it does not originate, which would corrupt the evaluation signal.
System Prompt: Question Verifier ””” TASK: The following question describes a visual Context. Verify if the cultural elements mentioned (Identitys, practices, symbols, customs) are grounded in actual {culture} cultural practices or facts: Question: ”{question}” Focus on whether the specific cultural elements mentioned are real and accurate for {culture} culture. OUTPUT FORMAT (Strict JSON, no extra text before or after): { "grounded": ”Yes” or ”No”, "evidence": ”1-sentence summary confirming or denying the cultural element is authentic” } ”””
C.6
Answer Generator
This stage evaluates each generated video against the verified question set from Stage. A vision-language model (Qwen3-VL-235B) is provided with a generated video and a single cultural question, and is asked to reason carefully before producing a binary Yes/No answer indicating whether the video is culturally faithful with respect to what the question asks. System Prompt: Answer Generator You are an expert cultural anthropologist and visual evaluator assessing the cultural faithfulness of a generated video. When answering each question, you MUST reason within <think> </think> tags following these steps: 1. Identify the visual evidence: Describe exactly what you observe in the video frames—specific Identitys, clothing details, spatial arrangements, architectural elements, lighting, and colors. 2. Assess cultural accuracy: Compare your observations against the culturally specific visual descriptions embedded in the question. Do not rely on implicit cultural knowledge—only evaluate what the question explicitly describes. 3. Evaluate temporal and physical coherence (for Behavior questions): Examine the sequence, duration, physics, and progression of movements across frames. Note whether Behaviors follow the temporal grounding specified in the question. 4. Check for stereotyping or inauthenticity: Flag if the video substitutes Western-centric defaults, hyper-exoticized elements, or generic representations in place of the specific cultural markers described in the question. After your reasoning, provide the final answer as either Yes or No. ”Yes” means the video is culturally faithful for what the question asks. ”No” means it fails or is culturally inaccurate. The final answer MUST BE put in a box. For example: Yes or No .
C.7
Cultural Uniqueness Classifier
This stage is a filtering step applied exclusively in the Geographical Constraint Removed ablation (RQ3). After country names are stripped from base prompts, not all remaining prompts are equally informative for probing implicit cultural knowledge. Some de-anchored prompts describe activities so universal — such as ”a family having dinner” or ”people watching television” — that no model, regardless of cultural competence, could reasonably be expected to generate a culturally specific output from them. Including such prompts in the ablation would dilute the signal and misattribute generic outputs as cultural failures.
System Prompt: Cultural Uniqueness Classifier You are a cultural analyst specializing in identifying whether a scene description is distinctly tied to a specific country or culture. Your task: Given a scene description (with the country name removed) and the target country, decide whether the described scene is culturally unique to that country. Definition of “culturally unique”: A prompt is culturally unique if it contains at least one element—such as a specific holiday, ritual, tradition, food, object, language term, landmark, or custom—that is distinctly and strongly associated with the given country. A casual observer familiar with world cultures would recognize it as belonging to that country. A prompt is NOT culturally unique if: • The scene could plausibly occur in any country (e.g., “a family having dinner”, “friends meeting at a cafe”) • The activity is common across many cultures (e.g., “people watching TV”, “a wedding ceremony”) • Only the time or season makes it distinct, not the cultural elements (e.g., “family watching fireworks on New Year’s Eve”) • The elements are broadly Western/universal rather than country-specific • The tradition or practice is shared across several neighboring countries without a country-specific distinguishing detail Examples: Country: Germany — Prompt: “Couple meeting at a sports club gathering” → is culturally unique: false → Sports club gatherings happen in virtually every country. Country: Japan — Prompt: “Tea ceremony with guests appreciating matcha” → is culturally unique: true → The Japanese tea ceremony (chado/chanoyu) is a codified cultural ritual unique to Japan. Country: Poland — Prompt: “Family sharing the opłatek wafer before Christmas Eve dinner” → is culturally unique: true → Sharing opłatek before the Wigilia supper is a distinctly Polish Catholic tradition. Apply the same careful reasoning to the input you receive.
D
Experiments
D.1
Model Specifications
Wan2.2 (Wan et al., 2025) We utilize the Wan2.2-T2V-A14B checkpoint for the text-to-video task. Generations are configured at a resolution of 1280 × 720 with a frame count of 49 to produce 5-second clips. The sampling process uses 20 steps with convert model dtype enabled and the T5 encoder running on CPU (--t5 cpu) to manage memory overhead. We generate a total of 2,943 videos (981 × 3 variations) for this model. LTX-2 (HaCohen et al., 2024) We employ a multi-component setup to ensure maximum output quality. This includes the ltx-2-19b-dev-fp8.safetensors base checkpoint, the gemma-3-12b-it-qat-q4 0-unquantized text encoder, and the ltx-2-19b-distilled-lora-384.safetensors for enhanced distillation. To achieve final resolution, we apply the ltx-2-spatial-upscaler-x2-1.0.safetensors spatial upsampler. Similar to Wan2.2, this model is evaluated across all 2,943 prompt instances. Veo 3.1 Fast (Google DeepMind) We use the veo-3.1-fast-generate-001 model via the Google DeepMind API. Due to the high operational costs associated with the Veo API, we do not run the full suite of 981 prompts. Instead, we take a stratified sample across cultural categories to generate 294 videos. This allows us to maintain a statistically significant comparison with the other SOTA models while remaining within computational and budgetary constraints. In total, our generation pipeline produced 6,180 videos for evaluation. D.2
VideoScore across Models
VideoScore (He et al., 2024) is an automated evaluation metric designed to simulate fine-grained human feedback for generative video models. It calculates scores across: Visual Quality (spatial clarity, resolution, and aesthetic appeal), Temporal Consistency (the absence of flickering, warping, or sudden object morphing across frames),
Dynamic Degree (the presence of fluid and significant motion), Text-to-Video Alignment (how accurately the visual content matches the semantic intent of the text prompt), and Factual Consistency (adherence to physical laws and common-sense logic). In the following Tables 2, 3, and 4, we have reported average videoscore across all five dimensions for LTX-2 (HaCohen et al., 2024). For Wan2.2 (Wan et al., 2025), we have reported scores in Table 5. For Veo 3.1 (Google DeepMind), we have reported scores in Table 6, 7 and 8. D.3
CultureScore across Models
In the following, we categorize CultureScore across countries and categories for LTX-2(HaCohen et al., 2024), (Wan et al., 2025), and Veo 3.1 Fast (Google DeepMind). Wan 2.2 (Wan et al., 2025) consistently outperformed for Identity and Context, and Veo 3.1 Fast (Google DeepMind) for behavior. Table 9, 10 for overall comparison across prompting strategies. Table 11, 12 for lTX-2 (HaCohen et al., 2024), Table 13, 14, 15, and 16 for Wan 2.2 (Wan et al., 2025) and Veo 3.1 Fast (Google DeepMind) Further provide all scores across countries and categories.
Country
Category
Visual Quality
Temporal Consistency
Dynamic Degree
Text-to-Video Alignment
Factual Consistency
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
3.485 3.685 3.547 3.621 3.528
3.445 3.679 3.471 3.582 3.525
3.431 3.627 3.555 3.65 3.456
3.243 3.555 3.394 3.488 3.236
3.496 3.674 3.496 3.619 3.55
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
3.624 3.58 3.401 3.673 3.363
3.55 3.549 3.332 3.622 3.329
3.533 3.504 3.396 3.695 3.311
3.274 3.27 3.187 3.489 3.045
3.604 3.56 3.376 3.655 3.391
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
3.376 3.578 3.293 3.608 3.437
3.304 3.533 3.168 3.57 3.367
3.356 3.527 3.299 3.609 3.317
3.129 3.407 3.025 3.46 3.043
3.389 3.554 3.246 3.596 3.432
China China China China China
dates-of-significance etiquette family greetings religion
3.615 3.775 3.791 3.65 3.715
3.579 3.713 3.728 3.56 3.697
3.57 3.704 3.753 3.609 3.507
3.373 3.567 3.546 3.416 3.416
3.605 3.746 3.76 3.638 3.74
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
3.281 3.387 3.287 3.559 3.007
3.238 3.327 3.264 3.52 3.081
3.152 3.399 3.208 3.513 2.859
3.093 3.299 3.156 3.418 2.906
3.285 3.373 3.293 3.557 3.094
India India India India India
dates-of-significance etiquette family greetings religion
3.57 3.668 3.562 3.668 3.726
3.518 3.614 3.493 3.62 3.678
3.527 3.605 3.517 3.596 3.658
3.307 3.488 3.373 3.472 3.467
3.561 3.646 3.552 3.671 3.715
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
3.623 3.796 3.673 3.877 3.624
3.582 3.762 3.64 3.844 3.57
3.557 3.763 3.637 3.866 3.605
3.335 3.605 3.472 3.662 3.415
3.63 3.775 3.67 3.863 3.634
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
3.283 3.595 3.171 3.467 3.054
3.238 3.514 3.169 3.447 3.064
3.164 3.521 3.042 3.345 2.796
3.021 3.394 3.008 3.271 2.771
3.29 3.545 3.2 3.488 3.071
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
3.377 3.357 3.29 3.138 3.075
3.33 3.317 3.174 3.126 3.081
3.217 3.241 3.236 3.011 2.876
3.044 3.141 3.078 2.988 2.874
3.364 3.338 3.254 3.162 3.116
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
3.645 3.624 3.655 3.54
3.607 3.591 3.624 3.547
3.54 3.549 3.651 3.472
3.357 3.366 3.471 3.223
3.641 3.603 3.654 3.587
Table 2: Average VideoScore (LTX-2 (HaCohen et al., 2024)) for base prompts, grouped by country and category.
Country
Category
Visual Quality
Temporal Consistency
Dynamic Degree
Text-to-Video Alignment
Factual Consistency
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
3.579 3.386 3.263 3.465 3.646
3.547 3.322 3.259 3.431 3.628
3.586 3.416 3.271 3.493 3.625
3.434 3.291 3.216 3.385 3.451
3.583 3.367 3.297 3.462 3.657
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
3.576 3.302 3.268 3.205 3.541
3.516 3.274 3.207 3.203 3.519
3.544 3.21 3.269 3.25 3.523
3.338 3.193 3.164 3.194 3.404
3.563 3.299 3.231 3.222 3.555
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
3.365 3.252 3.287 3.442 3.516
3.304 3.228 3.225 3.387 3.487
3.379 3.18 3.329 3.454 3.514
3.186 3.15 3.223 3.306 3.366
3.363 3.209 3.277 3.433 3.531
China China China China China
dates-of-significance etiquette family greetings religion
3.543 3.773 3.703 3.504 3.482
3.524 3.727 3.688 3.47 3.49
3.459 3.741 3.654 3.413 3.281
3.328 3.57 3.463 3.346 3.272
3.53 3.743 3.699 3.525 3.503
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
3.281 3.035 2.995 3.03 3.311
3.21 2.994 2.974 3.018 3.341
3.277 3.003 3.008 3.04 3.142
3.088 2.989 3.005 3.039 3.117
3.259 3.001 2.97 3.052 3.341
India India India India India
dates-of-significance etiquette family greetings religion
3.448 3.56 3.586 3.411 3.387
3.421 3.508 3.51 3.406 3.405
3.34 3.483 3.52 3.271 3.233
3.252 3.377 3.352 3.264 3.234
3.451 3.553 3.583 3.442 3.437
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
3.453 3.669 3.277 3.518 3.459
3.432 3.641 3.239 3.501 3.479
3.414 3.623 3.259 3.518 3.366
3.279 3.528 3.138 3.405 3.341
3.456 3.656 3.243 3.541 3.483
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
3.352 3.432 3.332 3.126 3.203
3.328 3.383 3.304 3.147 3.217
3.229 3.381 3.31 2.875 2.992
3.176 3.305 3.22 3.011 3.039
3.353 3.413 3.34 3.169 3.223
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
3.35 3.21 3.299 3.241 3.328
3.298 3.184 3.214 3.166 3.318
3.312 3.189 3.338 3.283 3.204
3.164 3.142 3.238 3.156 3.157
3.341 3.189 3.284 3.247 3.325
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
3.546 3.137 3.454 3.508
3.521 3.138 3.461 3.511
3.506 3.097 3.429 3.434
3.33 3.032 3.34 3.195
3.566 3.165 3.491 3.559
Table 3: Average VideoScore (LTX-2(HaCohen et al., 2024)) for extended prompts, grouped by country and category.
Country
Category
Visual Quality
Temporal Consistency
Dynamic Degree
Text-to-Video Alignment
Factual Consistency
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
3.438 3.611 3.488 3.54 3.349
3.381 3.605 3.388 3.46 3.334
3.375 3.55 3.497 3.607 3.244
3.18 3.489 3.367 3.376 2.961
3.429 3.61 3.447 3.525 3.361
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
3.539 3.587 3.223 3.496 2.975
3.505 3.519 3.147 3.442 2.975
3.412 3.557 3.233 3.498 2.84
3.273 3.266 3.052 3.402 2.787
3.547 3.545 3.175 3.497 3.062
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
3.268 3.336 3.148 3.54 2.989
3.163 3.312 3.059 3.474 2.942
3.276 3.196 3.112 3.513 2.81
3.05 3.12 2.923 3.397 2.734
3.254 3.319 3.124 3.516 3.038
China China China China China
dates-of-significance etiquette family greetings religion
3.333 3.355 3.433 3.103 3.296
3.273 3.297 3.311 3.065 3.32
3.253 3.223 3.458 3.039 2.976
3.082 3.192 3.184 2.989 2.982
3.32 3.326 3.392 3.117 3.344
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
3.261 3.557 3.576 3.617 3.268
3.244 3.516 3.493 3.539 3.331
3.15 3.57 3.534 3.578 3.064
3.094 3.457 3.256 3.43 3.084
3.267 3.524 3.522 3.576 3.335
India India India India India
dates-of-significance etiquette family greetings religion
3.497 3.534 3.279 3.293 3.432
3.418 3.5 3.252 3.227 3.431
3.45 3.484 3.234 3.193 3.358
3.229 3.364 3.08 3.126 3.233
3.479 3.538 3.315 3.279 3.466
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
3.349 3.334 3.257 3.27 3.21
3.266 3.309 3.211 3.232 3.192
3.196 3.251 3.22 3.26 3.114
3.022 3.189 3.131 3.184 3.067
3.319 3.318 3.24 3.233 3.239
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
3.306 3.608 3.508 3.314 3.118
3.269 3.534 3.443 3.319 3.097
3.195 3.581 3.478 3.193 2.946
3.081 3.445 3.266 3.137 2.886
3.323 3.552 3.483 3.335 3.119
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
3.44 3.643 3.372 3.268 3.036
3.346 3.591 3.226 3.243 3.009
3.353 3.608 3.353 3.205 2.795
3.12 3.46 3.122 3.144 2.827
3.403 3.581 3.315 3.295 3.03
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
3.438 3.439 3.516 3.203
3.392 3.393 3.479 3.169
3.377 3.429 3.52 3.097
3.145 3.333 3.362 2.922
3.436 3.417 3.531 3.235
Table 4: Average VideoScore (LTX-2 (HaCohen et al., 2024)) for geographical constraint removed prompts, grouped by country and category.
Country
Category
Visual Quality
Temporal Consistency
Dynamic Degree
Text-to-Video Alignment
Factual Consistency
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
2.76 2.74 2.79 2.73 2.76
2.79 2.82 2.86 2.82 2.84
2.64 2.54 2.6 2.54 2.56
2.73 2.78 2.81 2.7 2.69
2.74 2.69 2.79 2.73 2.79
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
2.8 2.75 2.74 2.73 2.75
2.82 2.76 2.79 2.83 2.81
2.61 2.6 2.61 2.55 2.51
2.7 2.73 2.83 2.76 2.66
2.77 2.67 2.71 2.7 2.76
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
2.8 2.75 2.76 2.75 2.75
2.84 2.82 2.8 2.85 2.83
2.62 2.52 2.61 2.51 2.53
2.74 2.73 2.78 2.75 2.65
2.77 2.71 2.73 2.76 2.78
China China China China China
dates-of-significance etiquette family greetings religion
2.76 2.74 2.75 2.74 2.78
2.8 2.77 2.83 2.83 2.91
2.53 2.5 2.47 2.46 2.13
2.71 2.71 2.7 2.72 2.72
2.75 2.66 2.71 2.78 2.83
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
2.73 2.73 2.75 2.76 2.76
2.79 2.78 2.81 2.87 2.88
2.55 2.51 2.57 2.47 2.34
2.66 2.73 2.78 2.78 2.73
2.72 2.67 2.71 2.76 2.82
India India India India India
dates-of-significance etiquette family greetings religion
2.79 2.75 2.77 2.78 2.78
2.86 2.8 2.85 2.87 2.83
2.43 2.43 2.45 2.49 2.47
2.72 2.71 2.76 2.74 2.74
2.79 2.71 2.76 2.81 2.8
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
2.74 2.76 2.76 2.77 2.74
2.84 2.84 2.85 2.88 2.87
2.46 2.45 2.54 2.44 2.39
2.63 2.71 2.79 2.73 2.78
2.74 2.7 2.73 2.78 2.75
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
2.71 2.74 2.75 2.71 2.73
2.78 2.8 2.83 2.76 2.82
2.37 2.39 2.52 2.42 2.28
2.64 2.71 2.79 2.67 2.63
2.69 2.67 2.74 2.64 2.69
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
2.72 2.72 2.71 2.76 2.77
2.78 2.8 2.77 2.89 2.88
2.49 2.46 2.53 2.46 2.44
2.64 2.66 2.76 2.72 2.68
2.68 2.64 2.68 2.77 2.79
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
2.77 2.75 2.76 2.75
2.83 2.84 2.87 2.86
2.56 2.51 2.48 2.44
2.7 2.77 2.71 2.66
2.77 2.72 2.76 2.8
Table 5: Average Videoscore (Wan 2.2 (Wan et al., 2025)) for base prompt grouped by Country and Category.
Country
Category
Visual Quality
Temporal Consistency
Dynamic Degree
Text-to-Video Alignment
Factual Consistency
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
2.90 2.66 2.84 2.82 2.72
3.06 2.62 2.84 2.87 2.70
2.70 2.70 2.73 2.76 2.68
2.84 2.78 2.86 2.88 2.80
3.01 2.53 2.80 2.77 2.67
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
3.30 2.80 2.80 2.68 3.66
3.23 2.79 2.91 2.80 3.60
3.31 2.82 2.73 2.64 3.52
3.16 2.80 2.86 2.77 3.50
3.27 2.80 2.83 2.64 3.62
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
3.02 2.69 2.58 2.63 2.83
2.96 2.75 2.60 2.70 2.92
2.95 2.58 2.67 2.61 2.57
2.92 2.78 2.66 2.69 2.83
2.91 2.64 2.52 2.65 2.86
China China China China China
dates-of-significance etiquette family greetings religion
2.76 2.76 2.75 2.66 2.88
2.68 2.84 2.66 2.74 2.84
2.49 2.58 2.66 2.73 2.47
2.48 2.83 2.59 2.79 2.84
2.52 2.66 2.56 2.73 2.78
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
2.66 2.82 2.77 2.66 2.86
2.65 2.69 2.78 2.70 2.90
2.48 2.43 2.66 2.53 2.61
2.64 2.66 2.84 2.71 2.85
2.62 2.70 2.66 2.60 2.88
India India India India India
dates-of-significance etiquette family greetings religion
2.91 2.71 2.76 2.75 2.83
2.94 2.75 2.82 2.84 2.77
2.70 2.50 2.39 2.48 2.49
2.78 2.57 2.60 2.77 2.84
2.88 2.51 2.62 2.75 2.77
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
2.80 2.78 2.82 2.84 2.80
2.80 2.77 2.91 2.95 2.83
2.61 2.65 2.49 2.43 2.47
2.73 2.78 2.85 2.75 2.88
2.73 2.68 2.75 2.76 2.77
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
2.73 2.91 2.73 2.68 2.77
2.84 2.83 2.84 2.62 2.98
2.21 2.83 2.31 2.62 2.27
2.77 2.78 2.73 2.62 2.72
2.64 2.75 2.66 2.61 2.84
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
2.74 2.90 2.80 2.85 3.19
2.76 2.91 2.80 2.82 3.22
2.63 2.88 2.60 2.91 2.93
2.80 2.92 2.77 2.90 3.07
2.72 2.89 2.69 2.82 3.16
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
3.20 2.63 2.78 3.23
3.20 2.70 2.94 3.14
3.24 2.63 2.52 3.16
3.13 2.63 2.80 3.11
3.23 2.56 2.82 3.23
Table 6: Average Videoscore (Veo 3.1 Fast ((Google DeepMind))) for base prompt grouped by country and category.
Country
Category
Visual Quality
Temporal Consistency
Dynamic Degree
Text-to-Video Alignment
Factual Consistency
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
2.60 2.64 2.72 2.73 2.76
2.67 2.75 2.73 2.91 2.88
2.58 2.66 2.65 2.48 2.52
2.62 2.84 2.89 2.80 2.82
2.66 2.66 2.62 2.75 2.80
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
3.30 2.75 3.03 2.82 3.16
3.30 2.69 3.08 2.86 3.12
3.37 2.72 2.91 2.79 2.97
3.22 2.73 3.03 2.93 3.10
3.34 2.72 3.00 2.80 3.19
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
3.67 2.70 2.71 2.84 2.80
3.59 2.73 2.62 2.89 2.87
3.66 2.55 2.70 2.70 2.54
3.44 2.73 2.73 2.84 2.73
3.64 2.59 2.62 2.88 2.79
China China China China China
dates-of-significance etiquette family greetings religion
2.80 2.73 2.80 2.73 2.74
2.79 2.84 2.88 2.79 2.94
2.48 2.36 2.61 2.67 2.12
2.65 2.70 2.80 2.80 2.77
2.59 2.59 2.66 2.75 2.80
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
2.69 2.71 2.80 2.65 2.80
2.69 2.71 2.84 2.65 2.91
2.46 2.57 2.52 2.59 2.56
2.59 2.66 2.84 2.77 2.85
2.57 2.63 2.63 2.55 2.82
India India India India India
dates-of-significance etiquette family greetings religion
2.73 2.60 2.88 2.89 2.74
2.83 2.63 2.84 2.98 2.82
2.54 2.48 2.62 2.41 2.36
2.66 2.64 2.88 2.81 2.84
2.72 2.48 2.78 2.81 2.76
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
3.24 3.28 2.70 2.86 2.72
3.27 3.30 2.77 2.94 2.67
3.07 3.12 2.52 2.48 2.66
3.09 3.12 2.78 2.75 2.77
3.22 3.25 2.59 2.76 2.67
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
2.75 2.69 2.82 2.77 2.80
2.88 2.51 2.91 2.94 2.96
2.45 2.73 2.45 2.00 2.26
2.77 2.64 2.88 2.50 2.80
2.68 2.50 2.78 2.69 2.84
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
2.80 2.73 2.68 2.90 2.78
2.90 2.73 2.66 2.94 2.82
2.46 2.64 2.62 2.76 2.64
2.83 2.66 2.77 2.90 2.77
2.79 2.62 2.63 2.91 2.77
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
2.86 2.67 2.72 2.67
2.91 2.72 2.69 2.81
2.75 2.56 2.73 2.31
2.85 2.56 2.80 2.72
2.90 2.66 2.70 2.70
Table 7: Average VideoScore evaluation of Veo 3.1 Fast (Google DeepMind) for extended prompt generations across countries and categories.
Country
Category
Visual Quality
Temporal Consistency
Dynamic Degree
Text-to-Video Alignment
Factual Consistency
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
2.69 2.56 2.68 2.74 2.69
2.86 2.59 2.70 2.86 2.70
2.33 2.73 2.61 2.59 2.55
2.72 2.81 2.84 2.79 2.75
2.75 2.53 2.55 2.80 2.67
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
2.86 2.87 2.69 2.72 3.02
2.80 2.90 2.79 2.77 2.92
2.72 2.80 2.73 2.62 3.03
2.77 2.91 2.89 2.66 3.05
2.79 2.88 2.65 2.62 3.01
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
3.63 2.75 2.59 2.77 3.00
3.49 2.73 2.55 2.92 2.97
3.73 2.70 2.72 2.45 2.88
3.48 2.82 2.78 2.77 2.89
3.48 2.66 2.42 2.81 2.95
China China China China China
dates-of-significance etiquette family greetings religion
2.80 2.70 3.24 3.12 2.85
2.69 2.70 3.23 3.12 2.91
2.55 2.55 3.09 3.20 2.41
2.80 2.64 3.09 3.15 2.86
2.53 2.54 3.12 3.14 2.84
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
2.71 2.78 2.86 2.70 2.79
2.71 2.75 2.90 2.89 2.88
2.46 2.23 2.64 2.41 2.42
2.62 2.73 2.96 2.62 2.74
2.62 2.71 2.72 2.75 2.84
India India India India India
dates-of-significance etiquette family greetings religion
2.67 3.14 2.64 2.89 2.93
2.84 3.07 2.61 3.02 2.94
2.30 2.99 2.62 2.38 2.59
2.48 2.91 2.70 2.80 2.94
2.63 2.98 2.53 2.84 2.96
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
2.75 2.72 2.77 2.84 2.75
2.80 2.70 2.90 2.96 2.62
2.57 2.60 2.50 2.31 2.75
2.75 2.58 2.87 2.66 2.83
2.73 2.62 2.72 2.73 2.62
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
2.79 2.71 2.76 2.89 2.80
2.88 2.70 2.82 2.69 2.97
2.25 2.67 2.54 3.06 2.20
2.73 2.73 2.81 2.89 2.78
2.61 2.59 2.73 2.84 2.85
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
2.76 3.27 2.71 2.62 2.67
2.84 3.27 2.73 2.67 2.78
2.50 3.19 2.52 2.61 2.46
2.86 3.08 2.66 2.69 2.56
2.70 3.06 2.63 2.61 2.62
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
3.18 2.66 2.62 2.70
3.16 2.76 2.56 2.70
3.20 2.65 2.78 2.55
3.12 2.71 2.78 2.77
3.21 2.71 2.55 2.67
Table 8: Average VideoScore evaluation of Veo 3.1 Fast (Google DeepMind) for geographical Constrainst Removed Prompting.
Behavior
Context
Identity
Overall
Country
Base
Ext
Gain
Base
Ext
Gain
Base
Ext
Gain
Base
Ext
Gain
Brazil Canada Chile China Germany India Iran Japan Poland South Africa
33.3 29.2 27.5 29.5 32.4 29.3 25.8 39.9 31.8 21.8
46.8 54.1 44.5 46.8 53.8 46.2 38.5 55.2 48.6 34.7
+13.5 +24.9 +17.0 +17.3 +21.4 +16.9 +12.7 +15.3 +16.9 +12.9
44.3 27.3 26.4 42.1 39.7 41.4 36.0 41.0 33.8 30.3
63.8 54.2 50.6 71.2 64.6 66.2 61.2 74.5 59.8 49.4
+19.5 +26.9 +24.2 +29.1 +24.9 +24.8 +25.2 +33.5 +25.9 +19.2
21.4 19.0 20.9 31.1 33.0 34.2 27.7 27.6 27.8 23.2
39.5 40.5 35.3 55.3 47.8 55.8 46.2 46.9 43.8 34.9
+18.1 +21.5 +14.4 +24.2 +14.9 +21.6 +18.5 +19.3 +16.0 +11.6
32.7 24.9 24.8 34.0 35.0 34.6 29.7 36.2 31.1 25.0
49.7 49.3 43.2 57.4 55.2 55.6 48.3 58.6 50.5 39.5
+17.0 +24.4 +18.4 +23.4 +20.3 +20.9 +18.6 +22.4 +19.4 +14.5
Average
30.1
46.9
+16.9
36.2
61.6
+25.3
26.6
44.6
+18.0
30.8
50.7
+19.9
Table 9: Overall CultureScore (%) across countries and prompting strategies for LTX-2 (HaCohen et al., 2024). Ext implies extended prompt and Gain implies (%) gain from Base to Extended Prompt.
Behavior
Context
Identity
Overall
Country
Base
Ext
Gain
Base
Ext
Gain
Base
Ext
Gain
Base
Ext
Gain
Brazil Canada Chile China Germany India Iran Japan Poland South Africa
33.1 38.0 29.6 41.1 34.9 33.6 28.4 46.7 30.8 27.0
48.2 55.8 44.5 55.9 49.7 48.8 41.4 62.7 39.4 42.8
+15.1 +17.8 +14.9 +14.8 +14.8 +15.2 +13.0 +16.0 +8.6 +15.8
53.3 44.3 40.9 56.7 51.4 46.7 36.6 59.0 43.1 43.0
73.2 60.8 61.5 76.3 71.7 68.8 64.7 75.5 62.6 62.4
+19.9 +16.5 +20.6 +19.6 +20.3 +22.1 +28.1 +16.5 +19.5 +19.4
29.6 34.5 26.3 48.7 35.9 40.6 28.9 36.7 28.0 21.4
47.1 52.7 40.9 69.1 57.9 63.6 46.7 58.8 45.9 39.3
+17.5 +18.2 +14.6 +20.4 +22.0 +23.0 +17.8 +22.1 +17.9 +17.9
38.3 38.8 31.9 48.4 40.3 39.8 31.1 47.2 33.6 30.0
55.1 56.2 47.8 66.1 58.5 59.2 50.0 65.1 47.8 47.3
+16.8 +17.4 +15.9 +17.7 +18.2 +19.4 +18.9 +17.9 +14.2 +17.3
Average
34.3
48.9
+14.6
47.5
67.8
+20.2
33.1
52.2
+19.1
37.9
55.3
+17.4
Table 10: Overall CultureScore (%) across countries and prompting strategies for Wan 2.2 (Wan et al., 2025). Ext implies extended prompt and Gain implies (%) gain from Base to Extended Prompt.
Country
Category
Identity CultureScore (%)
Behavior CultureScore (%)
Context CultureScore (%)
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
24.5 21.2 18.4 19.5 20.4
35.6 27.6 36.2 44.7 23.6
38.6 51.7 51.1 53.7 32.0
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
18.8 14.5 16.7 36.7 15.4
25.0 33.3 28.9 50.0 17.6
16.9 27.3 40.0 33.3 30.6
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
29.6 19.7 9.3 30.3 8.3
28.7 28.0 12.2 50.0 20.6
19.3 35.0 23.1 43.8 18.2
China China China China China
dates-of-significance etiquette family greetings religion
24.5 44.3 30.8 31.6 20.9
23.9 35.4 30.9 39.4 21.7
43.4 42.9 39.6 51.4 32.6
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
24.0 40.0 33.3 51.4 27.3
23.8 45.0 32.3 44.7 21.9
30.6 53.4 42.1 53.1 25.8
India India India India India
dates-of-significance etiquette family greetings religion
29.7 34.2 34.0 55.0 29.6
30.9 40.5 13.3 36.8 19.4
36.3 50.0 38.5 54.5 35.8
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
26.2 26.0 20.0 36.8 36.1
24.5 28.6 18.8 35.1 22.9
32.6 38.0 36.4 41.7 34.4
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
19.5 41.7 24.4 26.7 26.0
30.9 45.6 36.4 66.7 25.5
35.6 46.3 36.4 48.8 40.8
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
28.4 37.5 14.3 34.5 24.4
31.5 36.0 19.6 58.8 22.9
38.1 43.8 33.3 25.0 18.6
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
17.3 24.7 34.2 25.5
19.6 25.6 40.6 8.0
29.0 30.7 35.9 27.7
26.8
30.3
36.6
Overall
Table 11: CultureScore for LTX-2 (HaCohen et al., 2024) base prompt by country and category.
Country
Category
Identity CultureScore (%)
Behavior CultureScore (%)
Context CultureScore (%)
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
42.6 43.9 34.7 34.1 37.0
46.7 41.4 55.3 57.4 36.4
54.5 79.3 64.4 63.4 62.0
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
39.6 45.5 31.5 56.7 35.9
51.1 50.0 51.1 93.3 35.3
52.8 67.3 42.0 73.3 38.9
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
29.6 49.3 14.8 54.5 36.1
52.1 42.0 26.5 68.4 26.5
42.0 61.7 44.2 65.6 48.5
China China China China China
dates-of-significance etiquette family greetings religion
50.9 65.9 55.8 39.5 58.1
36.3 56.6 45.5 60.6 43.5
68.9 79.8 62.5 62.9 76.7
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
41.3 55.4 48.7 51.4 48.5
53.5 65.0 41.9 60.5 37.5
68.4 72.4 52.6 59.4 58.1
India India India India India
dates-of-significance etiquette family greetings religion
56.2 52.1 51.1 72.5 51.9
44.9 50.0 44.4 50.0 43.5
63.7 81.2 59.0 75.8 52.8
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
36.4 57.1 42.2 47.4 55.6
40.2 37.7 35.4 54.1 22.9
57.9 73.2 54.5 47.2 68.8
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
43.4 53.6 48.9 46.7 42.0
50.0 61.1 52.3 79.6 34.5
60.6 87.8 68.2 81.4 81.6
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
50.5 39.3 34.7 48.3 40.0
51.9 40.0 48.2 67.6 37.5
59.0 56.2 73.8 64.3 48.8
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
22.7 44.7 47.4 35.3
30.8 45.1 37.5 24.0
42.0 50.7 59.0 55.3
45.0
47.0
62.1
Overall
Table 12: CultureScore for LTX-2 (HaCohen et al., 2024) (%) for extended prompt by country and category.
Country
Category
Identity CultureScore (%)
Behavior CultureScore (%)
Context CultureScore (%)
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
31.9 25.8 24.5 39.0 27.8
38.9 25.9 38.3 40.4 18.2
50.0 63.8 53.3 53.7 44.0
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
36.5 21.8 40.7 60.0 17.9
40.9 27.8 46.7 53.3 20.6
34.8 47.3 58.0 60.0 27.8
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
31.6 23.9 11.1 42.4 27.8
29.8 30.0 14.3 47.4 17.6
47.7 36.7 28.8 53.1 33.3
China China China China China
dates-of-significance etiquette family greetings religion
47.3 52.3 32.7 55.3 55.8
42.5 48.5 14.5 54.5 47.8
56.6 64.3 33.3 71.4 58.1
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
32.7 50.8 17.9 45.7 33.3
24.8 51.7 35.5 39.5 31.2
51.0 53.4 50.0 62.5 38.7
India India India India India
dates-of-significance etiquette family greetings religion
28.9 45.2 38.6 70.0 42.6
29.4 40.5 11.6 44.7 35.5
43.4 48.4 47.2 60.6 41.5
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
24.3 31.2 24.4 31.6 38.9
25.5 24.7 27.1 40.5 22.9
35.8 46.5 20.5 38.9 37.5
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
25.7 53.6 31.1 42.2 32.0
37.3 57.8 40.9 59.3 38.2
46.2 68.3 65.9 60.5 65.3
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
32.1 26.8 22.4 34.5 22.2
33.3 32.0 21.4 52.9 18.8
44.8 41.7 40.5 57.1 32.6
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
16.4 23.5 28.9 21.6
23.4 24.4 25.0 20.0
34.0 52.0 46.2 46.8
33.4
33.9
47.8
Overall
Table 13: CultureScore by country and category for Wan 2.2 (Wan et al., 2025) (Base Prompt).
Country
Category
Identity CultureScore (%)
Behavior CultureScore (%)
Context CultureScore (%)
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
43.6 48.5 44.9 39.0 51.9
52.2 34.5 36.2 66.0 43.6
72.7 74.1 84.4 80.5 62.0
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
52.1 43.6 55.6 70.0 43.6
63.6 50.0 57.8 76.7 41.2
57.3 69.1 60.0 80.0 52.8
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
46.9 43.7 13.0 63.6 41.7
45.7 54.0 34.7 55.3 32.4
60.2 71.7 51.9 81.2 60.6
China China China China China
dates-of-significance etiquette family greetings religion
68.2 71.6 50.0 68.4 79.1
47.8 62.6 40.0 75.8 54.3
80.2 86.9 64.6 62.9 81.4
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
44.2 73.8 48.7 62.9 54.5
45.5 68.3 45.2 55.3 37.5
75.5 74.1 71.1 75.0 61.3
India India India India India
dates-of-significance etiquette family greetings religion
62.5 56.2 70.2 77.5 53.7
48.5 55.4 28.9 57.9 54.8
65.5 73.4 74.4 78.8 62.3
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
36.4 46.8 48.9 52.6 52.8
44.1 37.7 39.6 51.4 42.9
68.4 83.1 45.5 55.6 59.4
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
46.9 66.7 48.9 68.9 60.0
54.5 72.2 54.5 88.9 43.6
70.2 90.2 72.7 67.4 77.6
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
48.6 35.7 42.9 41.4 46.7
48.1 32.0 33.9 44.1 35.4
67.6 60.4 66.7 64.3 58.1
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
33.6 41.2 44.7 39.2
46.7 51.2 43.8 32.0
60.0 64.0 74.4 59.6
51.3
49.8
69.3
Overall
Table 14: CultureScore (%) by country and category for Wan 2.2 (Wan et al., 2025) for Extended prompt.
Country
Category
Identity CultureScore (%)
Behavior CultureScore (%)
Context CultureScore (%)
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
33.3 0.0 50.0 0.0 0.0
66.7 28.6 80.0 71.4 66.7
50.0 66.7 66.7 50.0 60.0
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
0.0 50.0 66.7 16.7 50.0
0.0 50.0 80.0 25.0 100.0
0.0 20.0 50.0 33.3 16.7
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
50.0 83.3 33.3 40.0 28.6
33.3 33.3 0.0 66.7 66.7
0.0 80.0 20.0 60.0 66.7
China China China China China
dates-of-significance etiquette family greetings religion
33.3 83.3 0.0 28.6 28.6
16.7 60.0 0.0 28.6 66.7
33.3 60.0 50.0 83.3 42.9
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
28.6 50.0 33.3 66.7 20.0
28.6 33.3 40.0 42.9 0.0
50.0 100.0 50.0 16.7 33.3
India India India India India
dates-of-significance etiquette family greetings religion
16.7 33.3 14.3 33.3 50.0
0.0 50.0 20.0 50.0 42.9
66.7 33.3 20.0 25.0 16.7
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
14.3 66.7 16.7 50.0 57.1
16.7 14.3 42.9 80.0 33.3
33.3 66.7 66.7 50.0 66.7
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
33.3 0.0 0.0 33.3 33.3
16.7 0.0 57.1 62.5 16.7
50.0 66.7 28.6 33.3 33.3
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
16.7 0.0 0.0 0.0 16.7
50.0 0.0 14.3 71.4 12.5
16.7 50.0 0.0 66.7 0.0
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
16.7 16.7 16.7 71.4
0.0 16.7 60.0 16.7
14.3 0.0 40.0 33.3
30.0
35.7
41.0
Overall
Table 15: CultureScore across country for Veo Fast 3.1 (Google DeepMind) for Base Prompt across all countries and categories.
Country
Category
Identity CultureScore (%)
Behavior CultureScore (%)
Context CultureScore (%)
Brazil Brazil Brazil Brazil Brazil
dates-of-significance etiquette family greetings religion
50.0 14.3 50.0 50.0 0.0
100.0 14.3 83.3 71.4 33.3
83.3 50.0 66.7 83.3 40.0
Canada Canada Canada Canada Canada
dates-of-significance etiquette family greetings religion
0.0 66.7 50.0 66.7 50.0
57.1 85.7 57.1 71.4 50.0
57.1 50.0 28.6 83.3 33.3
Chile Chile Chile Chile Chile
dates-of-significance etiquette family greetings religion
50.0 66.7 33.3 60.0 42.9
71.4 42.9 16.7 62.5 50.0
20.0 60.0 16.7 40.0 50.0
China China China China China
dates-of-significance etiquette family greetings religion
33.3 83.3 50.0 28.6 57.1
16.7 42.9 42.9 71.4 57.1
83.3 100.0 60.0 83.3 100.0
Germany Germany Germany Germany Germany
dates-of-significance etiquette family greetings religion
28.6 50.0 50.0 83.3 0.0
28.6 16.7 50.0 85.7 0.0
50.0 60.0 33.3 66.7 50.0
India India India India India
dates-of-significance etiquette family greetings religion
33.3 66.7 71.4 66.7 50.0
42.9 25.0 71.4 50.0 57.1
50.0 83.3 66.7 75.0 16.7
Iran Iran Iran Iran Iran
dates-of-significance etiquette family greetings religion
28.6 16.7 33.3 66.7 71.4
16.7 14.3 57.1 71.4 66.7
66.7 83.3 50.0 50.0 83.3
Japan Japan Japan Japan Japan
dates-of-significance etiquette family greetings religion
16.7 33.3 33.3 33.3 33.3
16.7 33.3 85.7 87.5 42.9
66.7 100.0 57.1 83.3 50.0
Poland Poland Poland Poland Poland
dates-of-significance etiquette family greetings religion
50.0 50.0 28.6 16.7 50.0
66.7 42.9 71.4 75.0 37.5
50.0 60.0 57.1 66.7 57.1
South Africa South Africa South Africa South Africa
dates-of-significance etiquette greetings religion
42.9 33.3 16.7 28.6
37.5 50.0 71.4 28.6
42.9 33.3 33.3 66.7
42.3
51.5
59.2
Overall
Table 16: CultureScore across country for Veo 3 Fast 3.1 (Google DeepMind) for Extended Prompt across all countries and categories.