LingT2I: On the Limitations of Cross-Lingual Consistency in Multilingual Text-to-Image Generation

A benchmark covering 10 widely used languages with 33K prompts, designed to evaluate cross-lingual effects in both content generation and text rendering. Building on this benchmark, we conduct a comprehensive cross-lingual analysis, uncovering linguistic inequality and language-dependent trade-offs across evaluation dimensions.

Quick Start

LingT2I Benchmark

Load from 🤗 Hugging Face.

from datasets import load_dataset

ds_cg = load_dataset(
    "RISys-Lab/LingT2I",
    split="content_generation",
)
ds_tr = load_dataset(
    "RISys-Lab/LingT2I",
    split="text_rendering",
)

sample_cg = ds_cg[0]
sample_tr = ds_tr[0]

print(sample_cg["prompt"])
print(sample_cg["dimension"], sample_cg["lang"])

print(sample_tr["prompt"])
print(sample_tr["render_text"])
print(sample_tr["condition_image"])

The benchmark contains two tasks:

• content_generation evaluates cross-lingual consistency for multilingual text-to-image prompts.
• text_rendering evaluates multilingual rendering with render text, layout metadata, and an embedded condition image.

Content-generation scoring uses lingt2i/evaluation/metaclip2_score.py. Text-rendering evaluation uses lingt2i/evaluation/ocr.py.

Setup

Installation

conda create -n lingt2i python=3.10 -y
conda activate lingt2i
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

For VLM-based evaluation, deploy an OpenAI-compatible endpoint with vLLM:

pip install accelerate
pip install 'vllm>=0.7.2'

vllm serve Qwen/Qwen2.5-VL-7B-Instruct \
  --port 8000 \
  --device cuda \
  --host 0.0.0.0 \
  --dtype bfloat16 \
  --limit-mm-per-prompt image=5,video=5

Getting Started

Content Generation

Set up a YAML file in configs/:

name: "lingt2i-content"
dataset_name: "RISys-Lab/LingT2I"
split: "content_generation"
start_idx: 0
end_idx: 30000
output_dir: "outputs/content_generation"

generation:
  models: ["zimage"]

Run the shared content-generation pipeline:

python -m lingt2i.generation.content \
  --config configs/content_generation.yaml

For the multilingual FLUX adapter:

python -m lingt2i.generation.pea

Text Rendering Generation

Prompt-only text-rendering models:

python -m lingt2i.generation.flux --start_idx 0 --end_idx 10
python -m lingt2i.generation.qwen --start_idx 0 --end_idx 10
python -m lingt2i.generation.seedream --start_idx 0 --end_idx 10
python -m lingt2i.generation.nano --start_idx 0 --end_idx 10
python -m lingt2i.generation.imagen4 --start_idx 0 --end_idx 10

Placement-aware text-rendering models:

python -m lingt2i.generation.anytext --max_samples 10
python -m lingt2i.generation.anytext2
python -m lingt2i.generation.easytext --max_samples 10

AnyText, AnyText2, and EasyText implementations are kept under lingt2i/third_party/, while benchmark entry points remain under lingt2i/generation/.

Content Evaluation

Use MetaCLIP2 to evaluate multilingual content alignment:

python -m lingt2i.evaluation.metaclip2_score \
  --image_folder outputs/content_generation/zimage \
  --dataset_name RISys-Lab/LingT2I \
  --split content_generation \
  --out_csv results/metaclip2_zimage.csv

Text Rendering Evaluation

Run OCR and text-rendering metrics:

python -m lingt2i.evaluation.ocr \
  --model_path outputs/text_rendering/EasyText \
  --dataset_name RISys-Lab/LingT2I \
  --split text_rendering \
  --ocr_mode gemini \
  --use_position \
  --output_file results.json

The OCR output contains:

• Character-level normalized edit distance.
• Token-level normalized edit distance.
• Exact sentence accuracy.
• Word accuracy.
• A combined average score.

Additional evaluation modules include:

• demographic_bias.py for demographic representation.
• cultural_bias.py for culture-specific elements and bias.
• nsfw.py for multilingual safety.
• trig_score.py for multilingual TRIG dimension scoring.
• summary.py for compact per-language OCR summaries.

Repository Structure

• configs/: experiment and model-path configuration.
• lingt2i/data.py: Hugging Face and legacy JSON data loaders.
• lingt2i/generation/: content-generation and text-rendering entry points.
• lingt2i/evaluation/: content, OCR, bias, cultural, and safety evaluation.
• lingt2i/models/: the copied text-to-image model layer shared with TRIG.
• lingt2i/analysis/: dataset and result analysis helpers.
• lingt2i/tools/: data-preparation and X2I projection utilities.
• lingt2i/third_party/: AnyText, AnyText2, and EasyText implementations.
• assets/fonts/: multilingual font resources.

Acknowledgement

Many thanks to the great works in multilingual image generation, including FLUX, Qwen-Image, AnyText, AnyText2, and EasyText.

Citation

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

The LingT2I BibTeX entry will be added when the paper is released.

@misc{zhang2026limitationscrosslingualconsistencymultilingual,
      title={On the Limitations of Cross-Lingual Consistency in Multilingual Text-to-image Generation}, 
      author={Sicheng Zhang and Zhonghao Yan and Binzhu Xie and Shi Qiu and Muzammal Naseer and Naveed Akhtar and Mubarak Shah},
      year={2026},
      eprint={2608.11002},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.11002}, 
}