Yesianrohn/waterec-wordart-recognition

WATERec WordArt Recognition

This Space runs the WATERec model for WordArt-Oriented Scene Text Recognition, as described in the paper "Advancing WordArt-Oriented Scene Text Recognition: Datasets and Methods" (ECCV 2026). Demo built by Apolinário (@apolinario) 🤝 Hugging Face. Thanks for bringing WATERec to life on Spaces!

Usage

Upload a cropped word/line image containing artistic or stylized text, and the model will recognize the text content. The model uses a NaViT-like RoPE-ViT encoder with an autoregressive Transformer decoder, supporting arbitrary-shaped inputs (curved, vertical, multi-oriented text).

Model

• —Architecture: NaViT_ROPE encoder (12 layers, 384-dim, 6 heads) + NaARDecoder (2 layers, 12 heads)
• —Checkpoint: WATERec-RS.pth (90.40% on WordArt-Bench)
• —Character set: 94 tokens (digits, letters, common symbols)
• —Max text length: 25 characters

Links

• —Paper (arXiv)
• —WATER repo
• —Model code (OpenOCR-WATERec)
• —Model weights