The Western Default in Text to Image Synthesis

This direction is split into other directions: culturally-blind evaluation. Its last version stays readable and citable here.

Version 1 of 169 papers109 conceptsfrozen 2026-10-04

When you prompt a state-of-the-art text-to-image model with a neutral noun like "a house" or "a wedding," the model does not sample from the full distribution of human experience. Instead, it typically defaults to a suburban American home or a bride in a white dress. This phenomenon, often called the "Western-centric bias," is not merely a lack of variety; it is a structural collapse of the model's latent space toward a narrow set of cultural centroids.

Research across dozens of benchmarks reveals that as these models become more photorealistic, they often become less representative, reinforcing a "Global North" default that erases the visual reality of the majority of the world.

The Erasure of the Underspecified

The most pervasive form of cultural bias is the default bias. In Inspecting the Geographical Representativeness of Images from Text-to-Image Models, Abhipsa Basu and colleagues demonstrate that when prompts lack geographic context, models like DALL-E 2 and Stable Diffusion prioritize U.S. aesthetics. This suggests that the underlying training distributions treat Western cultural artifacts as the "standard" representation for general concepts.

This isn't just a matter of preference; it's a performance gap. Melissa Hall and the Meta FAIR team introduced the DIG In framework, which found that models consistently demonstrate lower realism and diversity for regions like Africa and West Asia. Interestingly, they identified a "diversity tax": as models are tuned for higher aesthetic quality and prompt consistency, they often converge on stereotypical representations, narrowing the represented worldview even further.

Fossilization and the Exoticism Trap

When models do attempt to represent non-Western cultures, they often fall into "cultural fossilization" or "exoticism."

  1. Fossilization: Zhixuan Liu and others in the CCUB study note that models often depict non-Western societies through historical or primitive lenses. A prompt for a "house in Kenya" might yield a mud hut rather than a modern apartment in Nairobi.
  2. Exoticism: Sourojit Ghosh documented how models "lock" individuals from the Global South into traditional attire. While a Westerner might be depicted in hiking gear or an apron, a person from an Indigenous community is frequently rendered in ceremonial dress regardless of the activity specified in the prompt.

This creates a "stereotypical pull." Akshita Jha and Sunipa Dev found in their ViSAGe dataset that generated images are three times more likely to depict stereotypical traits than real-world images, forcing generations toward clichés even when users explicitly try to counter them.

Regional Blur and Asymmetric Stylization

Technically, these failures manifest as a lack of granular knowledge. Zahra Bayramli and colleagues describe a "regional blur" effect, where specific cultural artifacts (like a Korean hangari) are substituted with generic regional styles or features from neighboring high-resource cultures.

This is often accompanied by asymmetric stylization. Arif Ahmad and the team behind the CUBE benchmark observed that while Western subjects benefit from high photographic accuracy, non-Western subjects are more likely to be rendered as caricatures or with structural inaccuracies. The model understands the concept of the culture but lacks the data to render its specific geometry correctly.

Unlocking Cultural POV

If the problem is a lack of data in the base model, how can we fix it? Researchers are pursuing several paths:

  • Iterative Refinement: Suchae Jeong proposed Culture-TRIP, which uses an LLM to retrieve external knowledge (like Wikipedia descriptions) to refine prompts with structured cultural criteria before generation.
  • Multilingual Priming: Mor Ventura and Roi Reichart discovered that models often have "latent" cultural knowledge that can be "unlocked" by adding characters or optimized "gibberish" from a target language to the prompt.
  • Specialized Adapters: Projects like CTA-Flux by Yue Gong and others use lightweight adapters to bake native linguistic and cultural understanding into existing high-quality models, shifting the underlying visual distribution without retraining the entire system.

The goal of this research direction is to move beyond "aesthetic" quality toward "cultural competence"—ensuring that the "thousand words" an image speaks are accurate to the heritage and lived reality of the people it depicts.

Papers in this version

Showing 60 of 69 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_the_western_default_in_text_to_image_syn_40613bf7,
  title = {The Western Default in Text to Image Synthesis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-western-default-in-text-to-image-synthesis/rdr_40613bf7f1b8492f9e5ca30ceb2f3a6b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_the_western_default_in_text_to_image_syn_40613bf7_v1,
  title = {The Western Default in Text to Image Synthesis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-western-default-in-text-to-image-synthesis/rdr_40613bf7f1b8492f9e5ca30ceb2f3a6b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}