Measuring Cross-Market Generative Ability of
Vision–Language Models via Movie Poster Transcreation

Youyuan Lin  ·  Yuan Li  ·  Yahan Yu  ·  Fei Cheng  ·  Shinya Nishida  ·  Chenhui Chu

Kyoto University

Cross-market image transcreation requires preserving movie identity while adapting to market-specific design preferences and multilingual typography — a challenge that goes far beyond simple translation. We introduce MPTc-Bench, a benchmark of 582 aligned poster pairs spanning 34 target markets, sourced from Douban, Eiga, and IMDb. We define two task variants: Surface (text-centric localisation) and Deep (preference-level style adaptation), and propose a two-stage planner–editor pipeline. Evaluation combines information-retention checks, LLM-as-a-judge aesthetic scoring, and objective visual similarity signals. Our experiments reveal a substantial gap between model outputs and human-crafted target-market posters — particularly in faithful text rendering — and show that Gemini 3 currently leads across both task variants.

MPTc-Bench

Carefully curated aligned poster pairs with rich metadata, filtered from over 4,500 cross-market candidates using perceptual hashing and GLOBE cultural clustering.

Global coverage: 34 target markets across Asia, Europe, North and South America. Market selection is informed by GLOBE cultural clustering to ensure diversity.

from datasets import load_dataset
# Download: https://github.com/minamotooRin/mptc-bench/tree/main/data
import json, urllib.request
for split in ["surface", "deep"]:
    url = f"https://raw.githubusercontent.com/minamotooRin/mptc-bench/main/data/mptcbench_{split}.jsonl"
    # urllib.request.urlretrieve(url, f"mptcbench_{split}.jsonl")

Task Design & Evaluation

Poster pairs are split into two task levels using perceptual hash (pHash) distance, capturing fundamentally different localisation challenges.

Surface (top) vs. Deep (bottom) examples: Surface tasks preserve composition; Deep tasks require full visual redesign.

THREE-DIMENSIONAL EVALUATION

Cite This Work

If you use MPTc-Bench in your research, please cite our paper.

@inproceedings{lin2025mptcbench,
  title     = {Measuring Cross-Market Generative Ability of
               Vision--Language Models via Movie Poster Transcreation},
  author    = {Lin, Youyuan and Li, Yuan and Yu, Yahan and
               Cheng, Fei and Nishida, Shinya and Chu, Chenhui},
  booktitle = {Proceedings of the 63rd Annual Meeting of the
               Association for Computational Linguistics (ACL)},
  year      = {2025},
  note      = {Under review}
}