README.md

---
license: apache-2.0
task_categories:
  - text-to-image
language:
  - en
  - zh
pretty_name: PolyText-Bench
size_categories:
  - n<1K
tags:
  - text-to-image
  - visual-text-rendering
  - benchmark
  - multimodal
  - textdart
  - scanner
dataset_info:
  features:
    - name: pair_id
      dtype: string
    - name: task_type
      dtype: string
    - name: language
      dtype: string
    - name: scene_category
      dtype: string
    - name: prompt
      dtype: string
    - name: target_list
      list: string
    - name: target
      dtype: string
  splits:
    - name: critical_en
      num_examples: 150
    - name: optional_en
      num_examples: 150
    - name: critical_zh
      num_examples: 150
    - name: optional_zh
      num_examples: 150
configs:
  - config_name: default
    data_files:
      - split: critical_en
        path: critical_en.jsonl
      - split: optional_en
        path: optional_en.jsonl
      - split: critical_zh
        path: critical_zh.jsonl
      - split: optional_zh
        path: optional_zh.jsonl
---

# PolyText-Bench

> [**Project Page**](https://sujiehu.github.io/textdart-scanner/)

PolyText-Bench is a bilingual benchmark for evaluating text accuracy and glyph fidelity
in text-to-image generation. It is introduced in
[Blame Errant Words, Not the Canvas: Defect-Aware Spatial Credit Assignment for Visual Text Generation](https://arxiv.org/abs/2610.20794).

This repository contains both the **600 benchmark prompts** and the scripts required for
image generation, TextDART deployment, resumable scoring, and metric aggregation.

## Benchmark Design

PolyText-Bench contains paired critical and optional prompts in English and Chinese:

- **Critical prompts** specify one or more required text targets. They measure Critical
  Semantic accuracy (**CS**) and Critical Glyph quality (**CG**).
- **Optional prompts** describe the same scenes without prescribing text. They measure
  Optional Glyph quality (**OG**) over text that the generator introduces voluntarily.

| Split | Language | Task | Prompts |
|---|---|---|---:|
| `critical_en` | English | Required target text | 150 |
| `optional_en` | English | No required target text | 150 |
| `critical_zh` | Chinese | Required target text | 150 |
| `optional_zh` | Chinese | No required target text | 150 |

Critical and optional prompts with the same `pair_id` describe matched scenes. The
provided evaluation protocol generates four images per prompt by default using
deterministic paired seeds.

## Repository Contents

```text
PolyText-Bench/
├── critical_en.jsonl
├── optional_en.jsonl
├── critical_zh.jsonl
├── optional_zh.jsonl
├── generate.py
├── deploy_textdart.sh
├── evaluate.py
├── summarize.py
├── requirements.txt
└── LICENSE
```

## Load the Prompts

```python
from datasets import load_dataset

dataset = load_dataset("SujieHu/PolyText-Bench", "default")
critical_en = dataset["critical_en"]
optional_en = dataset["optional_en"]
critical_zh = dataset["critical_zh"]
optional_zh = dataset["optional_zh"]
```

Each record contains:

- `pair_id`: shared identifier for a critical/optional scene pair
- `task_type`: `critical` or `optional`
- `language`: `en` or `zh`
- `scene_category`: visual scene category
- `prompt`: text-to-image generation prompt
- `target_list`: required rendered strings; empty for optional prompts
- `target`: space-joined convenience representation of `target_list`

## Installation

Create the TextDART environment using the official source repository:

```bash
conda create -n TextDART python=3.11 -y
conda activate TextDART
git clone https://github.com/WeChatCV/TextDART-SCANNER.git
cd TextDART-SCANNER/TextDART/train
bash install_all.sh
cd ../..
```

Then clone this Dataset repository and install its scoring dependency:

```bash
git clone https://huggingface.co/datasets/SujieHu/PolyText-Bench
cd PolyText-Bench
pip install -r requirements.txt
```

The evaluation code uses the official
[WeChatCV/TextDART-SCANNER](https://github.com/WeChatCV/TextDART-SCANNER)
implementation for TextDART parsing and target matching.

## Download and Serve TextDART

Passing the Hugging Face repository ID downloads the model
automatically:

```bash
bash deploy_textdart.sh SujieHu/TextDART-8B-Qwen3VL 1925 0
```

To download the checkpoint first and serve it from disk instead:

```bash
hf download SujieHu/TextDART-8B-Qwen3VL \
  --local-dir TextDART-8B-Qwen3VL
bash deploy_textdart.sh ./TextDART-8B-Qwen3VL 1925 0
```

## Generate Images

Generate English images with Qwen-Image:

```bash
torchrun --nproc_per_node=8 generate.py \
  --model-family qwen \
  --mode EN \
  --model_name qwen_image \
  --model_path Qwen/Qwen-Image-2512 \
  --gallery_output_dir outputs
```

For FLUX.1-dev, set `--model-family flux` and provide its model path. Add
`--ckpt_path /path/to/adapter` to evaluate a PEFT LoRA checkpoint. Generation defaults
to 1024 × 1024, 50 steps, and four repeats per prompt.

Generated files are written to `outputs/<model_name>_<LANG>/` as
`<pair_id>_<task_type>_<repeat>.png`. You may also evaluate externally generated images
if they follow this naming and resolution contract.

## Evaluate

With the TextDART server running:

```bash
python evaluate.py \
  --mode EN \
  --image-dir outputs/qwen_image_EN \
  --output-dir scores/qwen_image_EN \
  --server-urls http://127.0.0.1:1925
```

The evaluator validates every expected image, scores it with TextDART, writes each
successful result durably, and resumes from `eval_results.jsonl` after interruption.

Validate score completeness and write `summary.json` plus `scores.md`:

```bash
python summarize.py \
  --mode EN \
  --scores-dir scores/qwen_image_EN
```

## Metrics

- **CS (Critical Semantic accuracy):** mean semantic score over required text targets;
  unmatched targets receive zero.
- **CG (Critical Glyph quality):** mean glyph-quality score over required text targets;
  unmatched targets receive zero.
- **OG (Optional Glyph quality):** mean glyph quality over text regions detected in
  optional-prompt images; images without detected text are reported as undefined and
  excluded from the OG average.

All reported scores are normalized to `[0, 1]`; higher is better.

## Citation

If you find PolyText-Bench useful, please cite:

```bibtex
@article{hu2026textdartscanner,
  title={Blame Errant Words, Not the Canvas: Defect-Aware Spatial Credit Assignment for Visual Text Generation},
  author={Hu, Sujie and Zhang, Huaisong and Yang, Pujing and Wang, Qian and Li, Xiaomin and Leng, Xinhang and Li, Chen and Lyu, Jing and Li, Xiu and Pan, Ling},
  journal={arXiv preprint arXiv:2610.20794},
  year={2026}
}
```