Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Translation: Cross-Cultural Meme Transcreation with Vision-Language Models

Yuming Zhao, Peiyi Zhang, Oana Ignat

TL;DR

This work reframes meme adaptation as cross-cultural transcreation rather than translation, introducing a three-stage hybrid framework that preserves intent while culturally adapting visuals and language. It provides MemeXGen, the first bidirectional CN↔US meme corpus with 6,315 original–transcreated pairs per direction, enabling systematic cross-cultural evaluation. Through human and automated assessments, the study reveals consistent directional asymmetries (US→Chinese outperforming Chinese→US) and analyzes which humor and design elements transfer across cultures. The findings underscore the need for culturally diverse data and robust human-in-the-loop evaluation when extending multimodal generation into culturally sensitive creative tasks, and the work publicly releases data, prompts, and evaluation protocols to drive future research.

Abstract

Memes are a pervasive form of online communication, yet their cultural specificity poses significant challenges for cross-cultural adaptation. We study cross-cultural meme transcreation, a multimodal generation task that aims to preserve communicative intent and humor while adapting culture-specific references. We propose a hybrid transcreation framework based on vision-language models and introduce a large-scale bidirectional dataset of Chinese and US memes. Using both human judgments and automated evaluation, we analyze 6,315 meme pairs and assess transcreation quality across cultural directions. Our results show that current vision-language models can perform cross-cultural meme transcreation to a limited extent, but exhibit clear directional asymmetries: US-Chinese transcreation consistently achieves higher quality than Chinese-US. We further identify which aspects of humor and visual-textual design transfer across cultures and which remain challenging, and propose an evaluation framework for assessing cross-cultural multimodal generation. Our code and dataset are publicly available at https://github.com/AIM-SCU/MemeXGen.

Beyond Translation: Cross-Cultural Meme Transcreation with Vision-Language Models

TL;DR

This work reframes meme adaptation as cross-cultural transcreation rather than translation, introducing a three-stage hybrid framework that preserves intent while culturally adapting visuals and language. It provides MemeXGen, the first bidirectional CN↔US meme corpus with 6,315 original–transcreated pairs per direction, enabling systematic cross-cultural evaluation. Through human and automated assessments, the study reveals consistent directional asymmetries (US→Chinese outperforming Chinese→US) and analyzes which humor and design elements transfer across cultures. The findings underscore the need for culturally diverse data and robust human-in-the-loop evaluation when extending multimodal generation into culturally sensitive creative tasks, and the work publicly releases data, prompts, and evaluation protocols to drive future research.

Abstract

Memes are a pervasive form of online communication, yet their cultural specificity poses significant challenges for cross-cultural adaptation. We study cross-cultural meme transcreation, a multimodal generation task that aims to preserve communicative intent and humor while adapting culture-specific references. We propose a hybrid transcreation framework based on vision-language models and introduce a large-scale bidirectional dataset of Chinese and US memes. Using both human judgments and automated evaluation, we analyze 6,315 meme pairs and assess transcreation quality across cultural directions. Our results show that current vision-language models can perform cross-cultural meme transcreation to a limited extent, but exhibit clear directional asymmetries: US-Chinese transcreation consistently achieves higher quality than Chinese-US. We further identify which aspects of humor and visual-textual design transfer across cultures and which remain challenging, and propose an evaluation framework for assessing cross-cultural multimodal generation. Our code and dataset are publicly available at https://github.com/AIM-SCU/MemeXGen.
Paper Structure (41 sections, 5 figures, 9 tables)

This paper contains 41 sections, 5 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Cultural Gaps in AI.
  4. Meme Understanding and Analysis.
  5. Cross-Cultural Transcreation.
  6. Generative Vision–Language Models.
  7. Evaluation of Multimodal Generation.
  8. Hybrid Transcreation Framework
  9. Meme Transcreation Pipeline
  10. MemeXGen Dataset
  11. Data Sources
  12. Filtering and Dataset Composition
  13. Annotation and Evaluation Split
  14. Dataset Characteristics
  15. Evaluation Methodology
  16. ...and 26 more sections

Figures (5)

  • Figure 1: Examples of cultural differences in meme preferences across Chinese and US contexts. Cultural preferences shape humor and visual style, creating challenges for cross-cultural meme transcreation.
  • Figure 2: Overview of our three-stage meme transcreation pipeline. (1) A VLM analyzes the original/input meme, identifies cultural references and intent, and generates a culturally adapted caption. (2) A diffusion model produces a meme-style visual template aligned with the target culture. (3) A text overlay module assembles the output transcreated meme.
  • Figure 3: Qualitative examples of cross-cultural meme transcreation.Left: successful US$\rightarrow$Chinese adaptation preserving intent, humor, and cultural conventions. Right: failed Chinese$\rightarrow$US adaptation illustrating loss of intent and cultural mismatch.
  • Figure 4: Examples of successful Chinese$\rightarrow$US meme transcreations. Original (middle left — dog meme): "Honest smile" Original (bottom left — angry emoji meme): "You looking for a knuckle sandwich?"
  • Figure 5: Examples of successful US$\rightarrow$Chinese meme transcreations. Transcreated (top right — spongebob meme): "Dad thinks I'm up early studying, I'm really on my 14th straight gaming session" Transcreated (middle right — spongebob meme): "Kid: Mom, I'm playing a game Mom: I'm cooking Kid: Can you pause it? Mom: How dare you use my own teachings against me?" Transcreated (middle right — Spencer Wright meme): "Tiktok be spamming ads upfront, still not gonna pay for premium"