ECCV 2026
Paper (arXiv) | Demo | Model Code | Models | Dataset | Captions
TL;DR: We advance WordArt-oriented scene TExt Recognition (WATER) from both data and model perspectives. We construct WATER-S, a 2M-scale synthetic artistic text dataset, and propose WATERec, a strong STR baseline supporting arbitrary-shaped inputs. Our approach achieves 90.40% accuracy on WordArt-Bench, the first result exceeding 90%, surpassing both general-purpose and OCR-specialized VLMs by a large margin.
- [2026/06] Code and data are released.
- [2026/06] Paper is accepted by ECCV 2026.
- WATER-S: A 2M-scale synthetic artistic text dataset consisting of two complementary subsets:
- WATER-T (1M): Tool-rendered via our SynthWordArt engine with 11,250 artistic fonts
- WATER-Z (1M): Generated by combining Qwen3-VL prompt mining + Z-Image synthesis
- WATER-R: A carefully deduplicated real training set (3.2M) from Union14M-L, WordArt, and WAS-R
- WATERec: An STR baseline with NaViT-like encoder (RoPE) for arbitrary-shaped inputs + AR decoder
- 90.40% accuracy on WordArt-Bench — first to exceed 90%, outperforming HunyuanOCR (81.54%) and other VLMs
WATER/
├── README.md
├── assets/ # Figures for README
├── SynthWordArt/ # WATER-T: artistic text rendering engine
│ ├── README.md
├── prompts/ # WATER-Z: prompt mining pipeline
│ ├── caption_mining.py # Step 1: mine captions from artistic text images
│ └── fewshot_expansion.py # Step 2: expand prompts via few-shot generation
├── Z-Image/ # WATER-Z: image generation with Z-Image
│ └── gen_zimage.py # Multi-GPU parallel generation script
└── eval_vlm/ # VLM evaluation on WordArt-Bench
├── get_acc.py # Accuracy computation
├── get_wrong.py # Error case extraction
├── infer_qwen3.py # Qwen3-VL-8B
├── infer_intern.py # InternVL3.5-8B
├── infer_got.py # GOT-OCR2.0
├── infer_deepseekocr.py # DeepSeek-OCR-2
├── infer_paddleocrvl.py # PaddleOCR-VL
├── infer_paddleocr.py # PP-OCRv5
├── infer_hunyuanocr.py # HunyuanOCR
└── infer_nemotron.py # Nemotron-VL-8B
External Repositories:
| Component | Link | Description |
|---|---|---|
| WATERec Demo | HuggingFace | WordArt Recognition Demo |
| WATERec Code | OpenOCR-WATERec | Model training & inference (based on OpenOCR) |
| WATERec Models | HuggingFace | Model ckpt |
| WATER-Data | HuggingFace | WATER-S, WATER-R, WordArt-Bench |
| WATER-Z Captions | HuggingFace | 273K prompt templates for WATER-Z generation |
| artistic-fonts | HuggingFace | 112K artistic fonts |
Pipelines of WATER-T (top) and WATER-Z (bottom).
Architecture of WATERec: NaViT-like encoder with RoPE + AR decoder.
If you find this work useful, please cite:
@inproceedings{water2026eccv,
title = {Advancing WordArt-Oriented Scene Text Recognition: Datasets and Methods},
author = {Ye, Xingsong and Du, Yongkun and Zhang, Jiaxin and Zhang, Haojie and Sun, Chong and Li, Chen and Lyu, Jing and Chen, Zhineng},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2026}
}- OpenOCR — Model training framework
- UnionST — Synthetic engine reference
- Union14M — Real dataset curation
- WordArt — WordArt benchmark
- Z-Image — Image generation for WATER-Z
- Qwen3-VL — VLM for prompt mining
This project is released under the Apache 2.0 License.
Font License Disclaimer: The artistic fonts used in WATER-T are collected from open-source platforms under their respective licenses (OFL, Apache, Creative Commons, etc.). If any font violates its license terms, please contact us and we will remove it promptly.