TexTeller/README.md

📄 English | <a href="./assets/README_zh.md">中文</a>

<div align="center">
    <h1>
        <img src="./assets/fire.svg" width=60, height=60>
        𝚃𝚎𝚡𝚃𝚎𝚕𝚕𝚎𝚛
        <img src="./assets/fire.svg" width=60, height=60>
    </h1>

  [![](https://img.shields.io/badge/API-Docs-orange.svg?logo=read-the-docs)](https://oleehyo.github.io/TexTeller/)
  [![](https://img.shields.io/badge/Data-Texteller3.0-brightgreen.svg?logo=huggingface)](https://huggingface.co/datasets/OleehyO/latex-formulas-80M)
  [![](https://img.shields.io/badge/Weights-Texteller3.0-yellow.svg?logo=huggingface)](https://huggingface.co/OleehyO/TexTeller)
  [![](https://img.shields.io/badge/docker-pull-green.svg?logo=docker)](https://hub.docker.com/r/oleehyo/texteller)
  [![](https://img.shields.io/badge/License-Apache_2.0-blue.svg?logo=github)](https://opensource.org/licenses/Apache-2.0)

</div>

https://github.com/OleehyO/TexTeller/assets/56267907/532d1471-a72e-4960-9677-ec6c19db289f

TexTeller is an end-to-end formula recognition model, capable of converting images into corresponding LaTeX formulas.

TexTeller was trained with **80M image-formula pairs** (previous dataset can be obtained [here](https://huggingface.co/datasets/OleehyO/latex-formulas)), compared to [LaTeX-OCR](https://github.com/lukas-blecher/LaTeX-OCR) which used a 100K dataset, TexTeller has **stronger generalization abilities** and **higher accuracy**, covering most use cases.

>[!NOTE]
> If you would like to provide feedback or suggestions for this project, feel free to start a discussion in the [Discussions section](https://github.com/OleehyO/TexTeller/discussions).



---

<table>
<tr>
<td>

## 🔖 Table of Contents
- [Getting Started](#-getting-started)
- [Web Demo](#-web-demo)
- [Server](#-server)
- [Python API](#-python-api)
- [Formula Detection](#-formula-detection)
- [Training](#️️-training)

</td>
<td>

<div align="center">
  <figure>
    <img src="assets/cover.png" width="800">
    <figcaption>
      <p>Images that can be recognized by TexTeller</p>
    </figcaption>
  </figure>
  <div>
  </div>
</div>

</td>
</tr>
</table>

## 📮 Change Log

<!-- - [2025-08-15] We have published the [technical report](https://arxiv.org/abs/2508.09220) of TexTeller. The model evaluated on the Benchmark (which was trained from scratch and had its handwritten subset filtered based on the test set) is available at https://huggingface.co/OleehyO/TexTeller_en. **Please do not directly use the open-source version of TexTeller3.0 to reproduce the experimental results of handwritten formulas**, as this model includes the test sets of these benchmarks. -->

- [2025-08-15] We have open-sourced the [training dataset](https://huggingface.co/datasets/OleehyO/latex-formulas-80M) of TexTeller 3.0. Please note that the handwritten* subset of this dataset is collected from existing open-source handwritten datasets (including both training and test sets). If you need to use the handwritten* subset for your experimental ablation, please filter the test labels first.

- [2024-06-06] **TexTeller3.0 released!** The training data has been increased to **80M** (**10x more than** TexTeller2.0 and also improved in data diversity). TexTeller3.0's new features:

  - Support scanned image, handwritten formulas, English(Chinese) mixed formulas.

  - OCR abilities in both Chinese and English for printed images.

- [2024-05-02] Support **paragraph recognition**.

- [2024-04-12] **Formula detection model** released!

- [2024-03-25] TexTeller2.0 released! The training data for TexTeller2.0 has been increased to 7.5M (15x more than TexTeller1.0 and also improved in data quality). The trained TexTeller2.0 demonstrated **superior performance** in the test set, especially in recognizing rare symbols, complex multi-line formulas, and matrices.

  > [Here](./assets/test.pdf) are more test images and a horizontal comparison of various recognition models.

## 🚀 Getting Started

1. Install uv:

   ```bash
   pip install uv
   ```

2. Install the project's dependencies:

   ```bash
   uv pip install texteller
   ```

3. If your are using CUDA backend, you may need to install `onnxruntime-gpu`:

   ```bash
   uv pip install texteller[onnxruntime-gpu]
   ```

4. Run the following command to start inference:

   ```bash
   texteller inference "/path/to/image.{jpg,png}"
   ```

   > See `texteller inference --help` for more details

## 🌐 Web Demo

Run the following command:

```bash
texteller web
```

Enter `http://localhost:8501` in a browser to view the web demo.

> [!NOTE]
> Paragraph recognition cannot restore the structure of a document, it can only recognize its content.

## 🖥️ Server

We use [ray serve](https://github.com/ray-project/ray) to provide an API server for TexTeller. To start the server, run the following command:

```bash
texteller launch
```

| Parameter | Description |
| --------- | -------- |
| `-ckpt` | The path to the weights file,*default is TexTeller's pretrained weights*. |
| `-tknz` | The path to the tokenizer,*default is TexTeller's tokenizer*. |
| `-p` | The server's service port,*default is 8000*. |
| `--num-replicas` | The number of service replicas to run on the server,*default is 1 replica*. You can use more replicas to achieve greater throughput.|
| `--ncpu-per-replica` | The number of CPU cores used per service replica,*default is 1*.|
| `--ngpu-per-replica` | The number of GPUs used per service replica,*default is 1*. You can set this value between 0 and 1 to run multiple service replicas on one GPU to share the GPU, thereby improving GPU utilization. (Note, if --num_replicas is 2, --ngpu_per_replica is 0.7, then 2 GPUs must be available) |
| `--num-beams` | The number of beams for beam search,*default is 1*. |
| `--use-onnx` | Perform inference using Onnx Runtime, *disabled by default* |

To send requests to the server:

```python
# client_demo.py

import requests

server_url = "http://127.0.0.1:8000/predict"

img_path = "/path/to/your/image"
with open(img_path, 'rb') as img:
    files = {'img': img}
    response = requests.post(server_url, files=files)

print(response.text)
```

## 🐍 Python API

We provide several easy-to-use Python APIs for formula OCR scenarios. Please refer to our [documentation](https://oleehyo.github.io/TexTeller/) to learn about the corresponding API interfaces and usage.

## 🔍 Formula Detection

TexTeller's formula detection model is trained on 3,415 images of Chinese materials and 8,272 images from the [IBEM dataset](https://zenodo.org/records/4757865).

<div align="center">
    <img src="./assets/det_rec.png" width=250>
</div>

We provide a formula detection interface in the Python API. Please refer to our [API documentation](https://oleehyo.github.io/TexTeller/) for more details.

## 🏋️‍♂️ Training

Please setup your environment before training:

1. Install the dependencies for training:

   ```bash
   uv pip install texteller[train]
   ```

2. Clone the repository:

   ```bash
   git clone https://github.com/OleehyO/TexTeller.git
   ```

### Dataset

We provide an example dataset in the `examples/train_texteller/dataset/train` directory, you can place your own training data according to the format of the example dataset.

### Training the Model

In the `examples/train_texteller/` directory, run the following command:

   ```bash
   accelerate launch train.py
   ```

Training arguments can be adjusted in [`train_config.yaml`](./examples/train_texteller/train_config.yaml).

## 📅 Plans

- [X] ~~Train the model with a larger dataset~~
- [X] ~~Recognition of scanned images~~
- [X] ~~Support for English and Chinese scenarios~~
- [X] ~~Handwritten formulas support~~
- [ ] PDF document recognition
- [ ] Inference acceleration

## ⭐️ Stargazers over time

[![Stargazers over time](https://starchart.cc/OleehyO/TexTeller.svg?variant=adaptive)](https://starchart.cc/OleehyO/TexTeller)

## 👥 Contributors

<a href="https://github.com/OleehyO/TexTeller/graphs/contributors">
   <a href="https://github.com/OleehyO/TexTeller/graphs/contributors">
      <img src="https://contrib.rocks/image?repo=OleehyO/TexTeller" />
   </a>
</a>