Discover ZYM-PKU/UDiffText Open Source project by ZYM (@ZYM-PKU)

UDiffText: A Unified Framework for High-quality Text Synthesis in Arbitrary Images via Character-aware Diffusion Models

Our proposed UDiffText is capable of synthesizing accurate and harmonious text in either synthetic or real-word images, thus can be applied to tasks like scene text editing (a), arbitrary text generation (b) and accurate T2I generation (c)

📬 News

2023.12.11 Version 2.0 update (getting rid of trash codes🚮)
2023.12.3 Build Hugging Face demo
2023.12.1 Build Github project page
2023.11.30 Version 1.0 upload

🔨 Installation

Clone this repo:

git clone https://github.com/ZYM-PKU/UDiffText.git
cd UDiffText

Install required Python packages

conda create -n udiff python=3.11
conda activate udiff
pip install -r requirements.txt

Make the checkpoint directory and build the tree structure

mkdir ./checkpoints

checkpoints
├── AEs                    // AutoEncoder
├── encoders             
    ├── LabelEncoder       // Character-level encoder
    └── ViTSTR             // STR encoder
├── predictors             // STR model
├── pretrained             // Pretrained SD
└── ***.ckpt               // UDiffText checkpoint

💻 Training

Prepare your data

LAION-OCR

Create a data directory {your data root}/LAION-OCR in your disk and put your data in it. Then set the data_root field in ./configs/dataset/locr.yaml.
For the downloading and preprocessing of Laion-OCR dataset, please refer to TextDiffuser and our ./scripts/preprocess/laion_ocr_pre.ipynb.

ICDAR13

Create a data directory {your data root}/ICDAR13 in your disk and put your data in it. Then set the data_root field in ./configs/dataset/icd13.yaml.
Build the tree structure as below:

ICDAR13
├── train                  // training set
    ├── annos              // annotations
        ├── gt_x.txt
        ├── ...
    └── images             // images
        ├── img_x.jpg
        ├── ...
└── val                    // validation set
    ├── annos              // annotations
        ├── gt_img_x.txt
        ├── ...
    └── images             // images
        ├── img_x.jpg
        ├── ...

TextSeg

Create a data directory {your data root}/TextSeg in your disk and put your data in it. Then set the data_root field in ./configs/dataset/tsg.yaml.
Build the tree structure as below:

TextSeg
├── train                  // training set
    ├── annotation         // annotations
        ├── x_anno.json    // annotation json file
        ├── x_mask.png     // character-level mask
        ├── ...
    └── image              // images
        ├── x.jpg.jpg
        ├── ...
└── val                    // validation set
    ├── annotation         // annotations
        ├── x_anno.json    // annotation json file
        ├── x_mask.png     // character-level mask
        ├── ...
    └── image              // images
        ├── x.jpg
        ├── ...

SynthText

Create a data directory {your data root}/SynthText in your disk and put your data in it. Then set the data_root field in ./configs/dataset/st.yaml.
Build the tree structure as below:

SynthText
├── 1                      // part 1
    ├── ant+hill_1_0.jpg   // image
    ├── ant+hill_1_1.jpg
    ├── ...
├── 2                      // part 2
├── ...
└── gt.mat                 // annotation file

Train the character-level encoder

Set the parameters in ./configs/pretrain.yaml and run:

python pretrain.py

Train the UDiffText model

Download the pretrained model and put it in ./checkpoints/pretrained/. You can ignore the "Missing Key" or "Unexcepted Key" warning when loading the checkpoint.

Set the parameters in ./configs/train.yaml, especially the paths:

load_ckpt_path: ./checkpoints/pretrained/512-inpainting-ema.ckpt // Checkpoint of the pretrained SD
model_cfg_path: ./configs/train/textdesign_sd_2.yaml // UDiffText model config
dataset_cfg_path: ./configs/dataset/locr.yaml // Use the Laion-OCR dataset

and run:

python train.py

📏 Evaluation

Download our available checkpoints and put them in the corresponding directories in ./checkpoints.
Set the parameters in ./configs/test.yaml, especially the paths:

load_ckpt_path: "./checkpoints/***.ckpt"  // UDiffText checkpoint
model_cfg_path: "./configs/test/textdesign_sd_2.yaml"  // UDiffText model config
dataset_cfg_path: "./configs/dataset/locr.yaml"  // LAION-OCR dataset config

and run:

python test.py

🖼️ Demo

In order to run an interactive demo on your own machine, execute the code:

python demo.py

or try our online demo at hugging face:

🎉 Acknowledgement

Dataset: We sincerely thank the open-source large image-text dataset LAION-OCR with character-level segmentations provided by TextDiffuser.
Code & Model: We build our project based on the code repo of Stable Diffusion XL and leverage the pretrained checkpoint of Stable Diffusion 2.0.

🪬 Citation

@misc{zhao2023udifftext,
      title={UDiffText: A Unified Framework for High-quality Text Synthesis in Arbitrary Images via Character-aware Diffusion Models}, 
      author={Yiming Zhao and Zhouhui Lian},
      year={2023},
      eprint={2312.04884},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

ZYM-PKU/UDiffText

ZYM-PKU

Reviews

Repository Details