---
license: apache-2.0
tags:
  - image-super-resolution
  - scene-text
  - text-image-guided-super-resolution
  - cvpr-2026
---

# TiGeSR checkpoints

Official checkpoints for **TiGeSR / TIGER**: *Restore Text First, Enhance Image Later: Two-Stage Scene Text Image Super-Resolution with Glyph Structure Guidance* (CVPR 2026).

- Paper: [CVPR Open Access](https://openaccess.thecvf.com/content/CVPR2026/html/Luo_Restore_Text_First_Enhance_Image_Later_Two-Stage_Scene_Text_Image_CVPR_2026_paper.html)
- Code: [OpenVeraTeam/TiGeSR](https://github.com/OpenVeraTeam/TiGeSR)
- Dataset: [UZ-ST](https://huggingface.co/datasets/mxluocv/UZ-ST)

## Checkpoint layout

```text
uzst/
├── stage1/
│   ├── model.pth
│   └── encoder.pth
└── stage2/
    └── model.pkl

realce/
├── stage1/
│   ├── model.pth
│   └── encoder.pth
└── stage2/
    └── model.pkl
```

`stage1` contains the glyph restoration model and text encoder. `stage2` contains the image-enhancement ControlNet checkpoint. Use the two checkpoints from the same dataset folder together.

## Download

```bash
hf download mxluocv/TiGeSR --local-dir weights
```

For inference commands and data preparation, see the code repository README.

## License

The code is released under Apache-2.0. Please consult the code repository and the respective base-model licenses for use of the checkpoints.