---
license: apache-2.0
language:
- zh
- en
- ja
- ko
- fr
- de
- es
- ru
- ar
- hi
- th
- el
library_name: pytorch
tags:
- ocr
- text-detection
- text-recognition
- paddleocr
- pp-ocrv5
- multilingual
- svtr
- db
pipeline_tag: image-to-text
---

# PP-OCRv5 PyTorch Model Zoo

PyTorch weights (**safetensors** format) for the full **PP-OCRv5** family, converted bit-exactly from the official PaddlePaddle `.pdparams` dynamic-graph weights — inference outputs are **identical to the original PaddleOCR** down to float32 precision.

- **Text Detection**: 2 models (mobile / server)
- **Text Recognition (base)**: 2 models covering Simplified Chinese / Traditional Chinese / English / Japanese
- **Text Recognition (multilingual)**: 11 models covering **100+ languages** (Korean, French, German, Russian, Arabic, Devanagari, Thai, Greek, Tamil, Telugu, etc.)

> This repo contains **weights + configs + dictionaries only**, not inference code. For inference, use [PaddleOCR2Pytorch](https://github.com/frotms/PaddleOCR2Pytorch), or follow the "Custom Python Inference" section below.
>
> Also available: [README_zh.md](./README_zh.md) (中文版).

---

## Repository Layout

```
.
├── README.md / README_zh.md
├── LICENSE                                                 # Apache 2.0
├── config.json                                             # Repo metadata + model index
│
├── ptocr_v5_*.safetensors                                  # 15 PP-OCRv5 weights at root (stable URLs)
├── ptocr_v5_server_{det,rec}.pth                           # Legacy pth copies of V5 server (kept)
│
├── configs/
│   ├── det/PP-OCRv5/
│   │   ├── PP-OCRv5_mobile_det.yml
│   │   └── PP-OCRv5_server_det.yml
│   └── rec/PP-OCRv5/
│       ├── PP-OCRv5_mobile_rec.yml                         # zh / zh-Hant / en / ja
│       ├── PP-OCRv5_server_rec.yml
│       └── multi_language/                                 # 11 multilingual rec yamls
│           ├── en_PP-OCRv5_mobile_rec.yaml
│           ├── korean_PP-OCRv5_mobile_rec.yml
│           ├── latin_PP-OCRv5_mobile_rec.yml               # French / German / Spanish / ...  (40+ Latin-script)
│           ├── eslav_PP-OCRv5_mobile_rec.yml               # Russian / Belarusian / Ukrainian
│           ├── cyrillic_PP-OCRv5_mobile_rec.yaml           # 33 Cyrillic-script languages
│           ├── arabic_PP-OCRv5_mobile_rec.yaml             # Arabic / Persian / Uyghur / Urdu / ...
│           ├── devanagari_PP-OCRv5_mobile_rec.yaml         # Hindi / Marathi / Nepali / Sanskrit / ...
│           ├── th_PP-OCRv5_mobile_rec.yaml                 # Thai
│           ├── el_PP-OCRv5_mobile_rec.yaml                 # Greek
│           ├── ta_PP-OCRv5_mobile_rec.yaml                 # Tamil
│           └── te_PP-OCRv5_mobile_rec.yaml                 # Telugu
│
├── dicts/                                                  # Character set dictionaries (required for rec)
│   ├── ppocrv5_dict.txt                                    # base (zh / zh-Hant / en / ja)
│   └── ppocrv5_<lang>_dict.txt                             # 11 multilingual dicts
│
└── legacy/                                                 # Older PP-OCR v2/v3/v4 weights (kept for back-compat)
    ├── ch_ptocr_mobile_v2.0_cls_infer.pth
    ├── ch_ptocr_v4_det_infer.pth
    ├── ch_ptocr_v4_rec_infer.pth
    ├── en_ptocr_v3_det_infer.pth
    └── en_ptocr_v4_rec_infer.pth
```

> All rec yamls use relative `character_dict_path: ./dicts/...`. After `git clone` or `snapshot_download`, paths resolve correctly with **no modification required**.

---

## Model Catalog

### Text Detection

| Weight | Config | Use case | Size |
|---|---|---|---|
| `ptocr_v5_mobile_det.safetensors` | `configs/det/PP-OCRv5/PP-OCRv5_mobile_det.yml` | Mobile / CPU-friendly | ~14 MB |
| `ptocr_v5_server_det.safetensors` | `configs/det/PP-OCRv5/PP-OCRv5_server_det.yml` | Server / high-accuracy | ~101 MB |

### Text Recognition (Base)

| Weight | Config | Languages | Size |
|---|---|---|---|
| `ptocr_v5_mobile_rec.safetensors` | `configs/rec/PP-OCRv5/PP-OCRv5_mobile_rec.yml` | Simplified / Traditional Chinese, English, Japanese | ~31 MB |
| `ptocr_v5_server_rec.safetensors` | `configs/rec/PP-OCRv5/PP-OCRv5_server_rec.yml` | same as above, higher accuracy | ~128 MB |

### Text Recognition (Multilingual)

All multilingual rec models share the same architecture (`SVTR_LCNet` + `PPLCNetV3`); they differ only by character dictionary. File size 23–28 MB each.

| Weight | Supported languages |
|---|---|
| `ptocr_v5_en_mobile_rec.safetensors` | English (dedicated model optimized for English-only scenarios) |
| `ptocr_v5_korean_mobile_rec.safetensors` | Korean, English |
| `ptocr_v5_latin_mobile_rec.safetensors` | French, German, Spanish, Italian, Portuguese, Dutch, Swedish, Danish, Norwegian, Finnish, Polish, Czech, Turkish, Vietnamese, ... (40+ Latin-script) |
| `ptocr_v5_eslav_mobile_rec.safetensors` | Russian, Belarusian, Ukrainian, English |
| `ptocr_v5_cyrillic_mobile_rec.safetensors` | 33 Cyrillic-script languages (Russian, Serbian-Cyrillic, Bulgarian, Mongolian, Kazakh, ...) |
| `ptocr_v5_arabic_mobile_rec.safetensors` | Arabic, Persian, Uyghur, Urdu, Pashto, Sindhi, ... |
| `ptocr_v5_devanagari_mobile_rec.safetensors` | 14 Devanagari-script languages (Hindi, Marathi, Nepali, Sanskrit, ...) |
| `ptocr_v5_th_mobile_rec.safetensors` | Thai, English |
| `ptocr_v5_el_mobile_rec.safetensors` | Greek, English |
| `ptocr_v5_ta_mobile_rec.safetensors` | Tamil, English |
| `ptocr_v5_te_mobile_rec.safetensors` | Telugu, English |

---

## Quick Start

### Download Weights

```python
from huggingface_hub import snapshot_download, hf_hub_download

# Option 1: download the whole repo (weights + configs + dicts + README)
repo_dir = snapshot_download(repo_id="JoyCN/PaddleOCR-Pytorch")
print("downloaded to:", repo_dir)

# Option 2: fetch a single weight file
weight_path = hf_hub_download(
    repo_id="JoyCN/PaddleOCR-Pytorch",
    filename="ptocr_v5_korean_mobile_rec.safetensors",
)
```

### Inference via PaddleOCR2Pytorch (Recommended)

```bash
# 1. clone the inference code repo
git clone https://github.com/frotms/PaddleOCR2Pytorch
cd PaddleOCR2Pytorch
pip install torch safetensors pyyaml shapely pyclipper opencv-python pillow scikit-image

# 2. Assume you ran snapshot_download above into /path/to/hf_repo
python tools/infer/predict_rec.py \
  --image_dir doc/imgs_words/korean/1.jpg \
  --rec_algorithm SVTR_LCNet \
  --rec_model_path /path/to/hf_repo/ptocr_v5_korean_mobile_rec.safetensors \
  --rec_yaml_path  /path/to/hf_repo/configs/rec/PP-OCRv5/multi_language/korean_PP-OCRv5_mobile_rec.yml \
  --rec_image_shape "3,48,320" \
  --rec_char_dict_path /path/to/hf_repo/dicts/ppocrv5_korean_dict.txt \
  --use_gpu False
```

> PaddleOCR2Pytorch `base_ocr_v20.py` auto-detects `.safetensors` vs `.pth` by extension (backward compatible).

### Custom Python Inference

A minimal skeleton showing how to load the weights and run a forward pass. You still need the network definitions from the PaddleOCR2Pytorch `pytorchocr/modeling/` package.

```python
import sys, numpy as np, cv2, torch, yaml
from safetensors.torch import load_file

# Requires https://github.com/frotms/PaddleOCR2Pytorch on PYTHONPATH
sys.path.insert(0, "/path/to/PaddleOCR2Pytorch")
from pytorchocr.modeling.architectures.base_model import BaseModel
from pytorchocr.postprocess import build_post_process

HF_REPO = "/path/to/hf_repo"   # the path returned by snapshot_download
yml_path    = f"{HF_REPO}/configs/rec/PP-OCRv5/multi_language/korean_PP-OCRv5_mobile_rec.yml"
weight_path = f"{HF_REPO}/ptocr_v5_korean_mobile_rec.safetensors"

# 1. load config + dictionary
with open(yml_path, encoding="utf-8") as f:
    cfg = yaml.safe_load(f)
dict_path = cfg["Global"]["character_dict_path"]          # './dicts/ppocrv5_korean_dict.txt'
dict_abs  = f"{HF_REPO}/{dict_path.lstrip('./')}"
with open(dict_abs, encoding="utf-8") as f:
    chars = [l.strip("\n\r") for l in f]
n_char = len(chars) + 2                                   # +1 blank, +1 space (if use_space_char)

# 2. build network + load weights (safetensors = zero-code-exec, mmap-fast)
cfg["Architecture"]["Head"]["out_channels_list"] = {
    "CTCLabelDecode": n_char,
    "SARLabelDecode": n_char + 2,
    "NRTRLabelDecode": n_char + 3,
}
net = BaseModel(cfg["Architecture"], out_channels=n_char)
net.load_state_dict(load_file(weight_path, device="cpu"))
net.eval()

# 3. preprocess (resize to [3, 48, 320], normalize to [-1, 1])
img = cv2.imread("input_word.jpg")
h, w = img.shape[:2]
ratio = w / h
tw = min(int(48 * ratio), 320)
img = cv2.resize(img, (tw, 48))
canvas = np.zeros((48, 320, 3), dtype=np.uint8)
canvas[:, :tw] = img
x = canvas.astype(np.float32).transpose(2, 0, 1) / 255.0
x = (x - 0.5) / 0.5
x = torch.from_numpy(x).unsqueeze(0)

# 4. forward + CTC decode
with torch.no_grad():
    logits = net(x)
post_op = build_post_process({
    "name": "CTCLabelDecode",
    "character_dict_path": dict_abs,
    "use_space_char": True,
})
result = post_op(logits)
print("prediction:", result)          # e.g. [('바탕으로', 0.9998)]
```

### Runtime Dependencies

```
torch        >= 1.13
safetensors  >= 0.4
numpy, pillow, opencv-python
pyyaml, shapely, pyclipper
scikit-image                # required by det post-processing
```

---

## Conversion & Verification

- **Source weights**: official PaddlePaddle `.pdparams` from  
  `https://paddle-model-ecology.bj.bcebos.com/paddlex/official_pretrained_model/`
- **Converter**: [PaddleOCR2Pytorch](https://github.com/frotms/PaddleOCR2Pytorch) — scripts `converter/ppocr_v5_det_converter.py` / `converter/ppocr_v5_rec_converter.py`
- **Verification**: end-to-end inference was run on macOS Apple Silicon (M-series) CPU; multilingual rec outputs are **bit-exact** with the original PaddleOCR `.pdparams` (float32 values match to 8 decimal places).

Sample inference results (CPU, < 0.7 s / image):

| Sample | Prediction | Confidence |
|---|---|---|
| Chinese `word_1.jpg` | 韩国小馆 | 0.99797755 |
| Korean `korean/1.jpg` | 바탕으로 | 0.99977183 |
| French `french/1.jpg` | de l'amendement, | 0.99656343 |
| Arabic `arabic/ar_1.jpg` | الكيصياوي | 0.68281130 |

---

## Legacy Files (`legacy/`)

Older PP-OCR (v2 / v3 / v4) checkpoints previously at the repo root have been **moved into `legacy/`** for clarity. They are still present and continue to work — just add the `legacy/` prefix to your path.

If you were previously using any of these URLs at the root:

```
legacy/ch_ptocr_mobile_v2.0_cls_infer.pth
legacy/ch_ptocr_v4_det_infer.pth
legacy/ch_ptocr_v4_rec_infer.pth
legacy/en_ptocr_v3_det_infer.pth
legacy/en_ptocr_v4_rec_infer.pth
```

The 15 PP-OCRv5 safetensors files **remain at the repo root** — their URLs did not change.

---

## License & Credits

- **License**: Apache License 2.0
- Weights originate from [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR) by the PaddlePaddle team (Apache 2.0).
- Converted with [PaddleOCR2Pytorch](https://github.com/frotms/PaddleOCR2Pytorch) (Apache 2.0).

If this repo helps you, please also star both of those original projects.

---

## Citation

```bibtex
@misc{pp_ocrv5_pytorch_joycn_2025,
  title        = {PP-OCRv5 PyTorch Model Zoo},
  author       = {JoyCN},
  howpublished = {\url{https://huggingface.co/JoyCN/PaddleOCR-Pytorch}},
  year         = {2025}
}
```