Current text-to-image (T2I) and vision-language models (VLMs) possess a "Global-North" default. When asked to generate a "house," models like DALL·E 2 and Stable Diffusion typically produce U.S.-style suburban homes; when asked to describe a Korean hanbok, they may settle for the generic label "traditional attire."
This cultural gap persists because our standard ways of measuring AI success—metrics like CLIPScore for alignment or FID for realism—are culturally blind. Recent research is shifting toward cultural competence, a framework that prioritizes the accurate representation of global heritage, artifacts, and social contexts over mere aesthetic polish.
The Failure of Standard Metrics
The primary obstacle to culturally inclusive AI is "metric drift." Standard automated metrics often fail to correlate with human judgment regarding cultural authenticity. In fact, as Sina Malakouti and colleagues demonstrate with the AHEaD framework in Culture in Action: Evaluating Text-to-Image Models through Social Activities, alignment metrics like CLIPScore can actually reward cultural stereotypes.
Because these metrics rely on keyword matching, they assign high scores to images that over-represent recognizable, stereotypical markers (like a caricature of a festival) while ignoring blatant hallucinations or contextual inaccuracies. Similarly, Zahra Bayramli found that Fréchet Inception Distance (FID) shows low or even negative correlation with human assessments of cultural authenticity in Diffusion Models Through a Global Lens: Are They Culturally Inclusive?. An image can be technically high-quality and photorealistic while remaining culturally nonsensical to a native observer.
Compositional Fragility and Geographic Bias
The cultural gap is not just a lack of data; it is an architectural and compositional failure. Muna Numan Said identifies a phenomenon called compositional fragility in Deconstructing Bias: A Multifaceted Framework for Diagnosing Cultural and Compositional Inequities in Text-to-Image Generative Models. Models often struggle to combine marginalized cultural elements correctly, such as placing a traditional non-Western ceremony inside a Western architectural setting.
Furthermore, there is a "diversity-polish trade-off." As models become more advanced and "polished," they often become less diverse. The GeoDiv framework reveals that newer models like FLUX.1 produce high-quality, "affluent" images but are significantly more prone to socio-economic stereotyping than their predecessors, defaulting to Western aesthetics to achieve a high-fidelity look.
Benchmarking Heritage and Folklore
To address these failures, researchers have introduced specialized benchmarks that move beyond general object recognition to test deep cultural knowledge:
- CUBE (CUltural BEnchmark): Arif Ahmad and team developed CUBE to evaluate how models handle artifacts from the Global South, revealing a tendency toward cultural homogenization.
- RusCode: In RusCode: Russian Cultural Code Benchmark for Text-to-Image Generation, Viacheslav Vasilev shows that Western models often fail to capture specific local identities, defaulting to Westernized approximations of Russian culture.
- MOSAIC-1.5k: Aman Chadha introduced the Cultural Awareness Score (CAS) in How Culturally Aware are Vision-Language Models? to quantify how well models recognize folklore, mythology, and religious symbols.
Toward Culturally-Aware Architectures
Fixing the cultural gap requires moving beyond simple fine-tuning. Researchers are exploring multi-stage pipelines and retrieval strategies to inject context into model outputs.
Youngsik Yun proposed the CIC (Culturally-aware Image Captioning) framework, which uses Visual Question Answering (VQA) to extract specific cultural details before using a Large Language Model to synthesize a description. This prevents the model from defaulting to generic labels. Similarly, Jiaang Li demonstrated in RAVENEA that Retrieval-Augmented Generation (RAG)—providing models with cultural reference documents—significantly boosts performance on culture-centric tasks, especially for smaller models.
By shifting the focus from "how good the image looks" to "how accurately the image represents the world," this research direction aims to ensure that generative AI serves as a tool for global expression rather than a mirror for a single cultural perspective.