Imagine typing the prompt "a photo of a person drinking tea" into a text-to-image generator. You receive an image of an individual in a modern Western kitchen. Now, imagine replacing just the letter "o" in "photo" with the Greek letter "ο" (omicron). To a human reader, the prompt looks completely identical. But to the underlying artificial intelligence, this single character swap changes everything: the generated image suddenly shifts to depict Mediterranean clothing, distinct ethnic features, or regional architectural styles.
This phenomenon is known as script-induced cultural bias. Recent research reveals that text-to-image models do not just interpret the semantic meaning of our words; they are highly sensitive to the underlying Unicode scripts and character encodings used to write them.
The Core Mechanism: Embedding Space Separation
To understand why a visually identical character can radically alter an image, we have to look inside the model's text encoder. State-of-the-art text-to-image models like Stable Diffusion and DALL-E 2 rely on multimodal text encoders (such as CLIP) to translate text prompts into continuous vector embeddings that guide the image generation process.
As Lukas Struppek et al. demonstrate in The Biased Artist: Exploiting Cultural Biases via Homoglyphs in Text-Guided Image Generation Models, these text encoders organize their embedding space by clustering characters based on their script origin. Because the models are trained on massive internet datasets where specific writing systems are heavily correlated with specific geographic and cultural contexts, the encoder mathematically links character types to cultural concepts.
When a user introduces a "homoglyph"—a character from a non-Latin script that looks identical to a Latin character—the text tokenizer breaks the word apart differently, and the encoder maps the prompt into a distinct region of the embedding space. This activation bypasses the literal meaning of the prompt and forces the generative process to follow implicit cultural directions based purely on character choice.
[Standard Prompt] -> "photo" ----> Latin Embedding Region ----> Standard Image
[Homoglyph Swap] -> "phοtο" ----> Greek Embedding Region ----> Culturally Biased Image
Script-Induced Cultural Hallucination
This vulnerability extends far beyond single-character manipulations. The bias is deeply embedded in how models process entirely different native writing systems. When a prompt is written in a non-Latin script, the model frequently prioritizes the orthography (the visual style of the script) over the actual semantics (the literal meaning of the text).
In Navigating Text-to-Image Generative Bias Across Indic Languages, Surbhi Mittal et al. introduced a comprehensive benchmark called IndicTTI to evaluate how text-to-image generators handle 30 different Indic languages. The authors discovered a pervasive issue they termed "script-induced cultural hallucination."
When mundane, culturally neutral prompts are written in native Indic scripts, the models regularly fail to generate the literal object requested. Instead, they default to narrow cultural archetypes, injecting religious imagery, traditional attire, or stereotypical festivals into the output. The script itself acts as a massive visual trigger that overrules the text's actual instructions, exposing a stark performance gap between English prompts and non-Western writing systems.
Neutralizing the Bias via Embedding Alignment
Because these biases are baked into the text encoder's representation space, resolving them requires making the encoder robust against character-level variations without destroying its ability to understand language.
To address this, Lukas Struppek et al. proposed a mitigation strategy in Exploiting Cultural Biases via Homoglyphs in Text-to-Image Synthesis. Instead of retraining an entire text-to-image model from scratch—which is computationally prohibitive—the authors developed a "homoglyph unlearning" method using a teacher-student fine-tuning framework.
During this fine-tuning process, the student text encoder is exposed to prompts containing homoglyphs and non-Latin script variations. The training objective forces the student model's output embeddings to mathematically align with the embeddings generated by a clean, standard Latin prompt processed by the teacher model. By pulling these disparate script regions together into a shared semantic space, the model learns that character-level variations should not alter the core meaning of the text. This effectively neutralizes the script-induced bias, preventing malicious or accidental stereotype propagation while fully preserving the model's overall image quality.
Key Foundational Work
Exploiting Cultural Biases via Homoglyphs inText-to-Image Synthesis (Abstract Reprint) (2024) — contributed 6 concept elements to this direction.
Exploiting Cultural Biases via Homoglyphs in Text-to-Image Synthesis (2024) — contributed 6 concept elements to this direction.
Navigating Text-to-Image Generative Bias Across Indic Languages (2024) — contributed 4 concept elements to this direction.