Where Culture Fades: Revealing the Cultural Gap in Text-to-Image Generation

TL;DR

Multilingual text-to-image models often generate culturally neutral or English-biased images under prompts in multiple languages. The authors show that culture-related knowledge is present but under-activated and introduce CultureBench, a probing framework, and two lightweight alignment strategies—inference-time cultural activation and layer-targeted enhancement—to boost cultural consistency. They demonstrate significant improvements on CultureBench in cultural attribution and visual fidelity, with results supported by automatic metrics and human evaluation. This work provides a practical path toward inclusive, culturally grounded image synthesis and offers a diagnostic benchmark for cross-cultural T2I systems.

Abstract

Multilingual text-to-image (T2I) models have advanced rapidly in terms of visual realism and semantic alignment, and are now widely utilized. Yet outputs vary across cultural contexts: because language carries cultural connotations, images synthesized from multilingual prompts should preserve cross-lingual cultural consistency. We conduct a comprehensive analysis showing that current T2I models often produce culturally neutral or English-biased results under multilingual prompts. Analyses of two representative models indicate that the issue stems not from missing cultural knowledge but from insufficient activation of culture-related representations. We propose a probing method that localizes culture-sensitive signals to a small set of neurons in a few fixed layers. Guided by this finding, we introduce two complementary alignment strategies: (1) inference-time cultural activation that amplifies the identified neurons without backbone fine-tuned; and (2) layer-targeted cultural enhancement that updates only culturally relevant layers. Experiments on our CultureBench demonstrate consistent improvements over strong baselines in cultural consistency while preserving fidelity and diversity.

Figures (19)

• Figure 1: Cultural alignment in local languages. (a) LLMs/recommenders keep cultural consistency, but T2I models falter with "noun-only’’ prompts. (b) Adding a "culture-style modifier + noun’’ restores consistency.
• Figure 2: Overview of the CultureBench pipeline. First, manually collect and rigorously quality-control datasets from 15 linguistic regions; annotate “culture-style modifier noun" captions using GPT5-Nano chatgpt5 and, through human annotation, “noun-only" captions; convert annotated content into local languages via translation tools, supplemented by manual review.
• Figure 3: Data distribution of the proposed CultureBench dataset across 15 languages. The dataset is divided into train, test, and neuron-detection subsets with a ratio of 7:2:1.
• Figure 4: Verify the hypothesis. Within the CultureBench test subset, performances under “culture-style modifier + noun” and “noun-only” prompt conditions are compared. Quantitative evaluation is conducted using CultureVQA.
• Figure 5: Methods for Neuronal Detection. (a) By comparing attention allocation between cultural-style modifiers and nouns across text-encoder layers, the layer with the largest divergence is designated as the culturally sensitive layer. (b) At this layer, features from the “culture-style modifier + noun’’ and “noun-only’’ prompts are fed into an SAE cunningham2023sparse to obtain sparse features, revealing neurons with heightened sensitivity to cultural cues.