Jiabao Ji¹, Guanhua Zhang¹, Zhaowen Wang², Bairu Hou¹, Zhifei Zhang², Brian Price², Shiyu Chang^1
1UC, Santa Barbara, ²Adobe Research

This is the official implementation of the paper "Improving Diffusion Models for Scene Text Editing with Dual Encoders" [Arxiv].

In this work, we propose a novel Diffusion-based Scene Text Editing (DiffSTE) framework, which is able to edit scene text into different font styles and colors following given text instruction. Specifically, we propose to improve pre-trained diffusion models with a dual encoder design, which includes a character encoder for better text legibility and an instruction encoder for better style control. We then utilize an instruction tuning framework to train our model learn the mapping from the text instruction to the corresponding image with either the specified style or the style of the surrounding texts in the background. Such a training method further brings our model the zero-shot generalization ability to the following three scenarios: generating text with unseen font variation, e.g. italic and bold, mixing different fonts to construct a new font, and using more relaxed forms of natural language as the instructions to guide the generation task.

Assuming conda has already installed, you could use the following commands to try our model for a quick start.

conda create -n diffste python=3.8
conda activate diffste
pip install -r requirements.txt
gdown https://drive.google.com/uc?id=1fc0RKGWo6MPSJIZNIA_UweTOPai64S9f
python generate.py --ckpt_path diffste.ckpt --in_image examples/sample0.png --in_mask examples/mask0.png --text wizards --out_dir ./

Build the environment with the following command:

conda create -n diffste python=3.8
conda activate diffste
pip install -r requirements.txt

Our pretrained model can be downloaded from here.

Run following command to edit scene text. The mask file indicates the region where the generated text locates.

python generate.py --ckpt_path ${model_path} --in_image examples/sample0.png --in_mask examples/mask0.png --text wizards --out_dir ${output_dir}

You should be able to get a similar result:

Specify the font and color of the generated text by adding --font and --color arguments.

python generate.py --ckpt_path ${model_path} --in_image examples/sample1.png --in_mask examples/mask1.png --text five --font Courgette --color red --out_dir ${output_dir}

You should be able to get a similar result:

Specify the text style with a natural language instruction.

python generate.py --ckpt_path ${model_path} --in_image examples/sample2.png --in_mask examples/mask2.png --text STAFF --instruction "The word \"STAFF\" is colored in a delicate, ladylike shade of lilac"" --out_dir ${output_dir}

You should be able to get a similar result:

Generate text with unseen font variation, e.g. italic and bold. Notice that NovaMono font has no italic and bold version from google-fonts library.

python generate.py --ckpt_path ${model_path} --in_image examples/sample3.png --in_mask examples/mask3.png --text STATION --font NovaMono --out_dir ${output_dir}
python generate.py --ckpt_path ${model_path} --in_image examples/sample3.png --in_mask examples/mask3.png --text STATION --font NovaMono-Italic --out_dir ${output_dir}
python generate.py --ckpt_path ${model_path} --in_image examples/sample3.png --in_mask examples/mask3.png --text STATION --font NovaMono-Bold --out_dir ${output_dir}
python generate.py --ckpt_path ${model_path} --in_image examples/sample3.png --in_mask examples/mask3.png --text STATION --font NovaMono-BoldItalic --out_dir ${output_dir}

You should be able to get similar results:

Mix two different font styles.

python generate.py --ckpt_path ${model_path} --in_image examples/sample4.png --text Reload --font Allura --out_dir ${output_dir}
python generate.py --ckpt_path ${model_path} --in_image examples/sample4.png --text Reload --font Mohave --out_dir ${output_dir}
python generate.py --ckpt_path ${model_path} --in_image examples/sample4.png --text Reload --font "Allura and Mohave" --out_dir ${output_dir}

You should be able to get similar results:

You can train the model on a combination of real world scene text data and synthetic scene text data.

1. Download real world dataset:

sh scripts/down_data.sh

2. Generate synthetic dataset:

pip install -r synthgenerator/requirements.txt
sh scripts/gen_synth.sh

Notice that you may need to first download fonts from google fonts library, we include a list of font names for our released model in synthgenerator/resources/100fonts and background images from SynthText Project.

The donwloaded real world data and synthetic data will be in folder data/ocr-dataset.

The main training script is train.py. You can train the model by running

python train.py --base ${config_paths} --stage fit --name ${run_name} --project ${project_name} --base_logdir ${log_directory}

Logs and model will be saved in ${log_directory}/${project_name}/${time}_${run_name}. An example config file is in configs folder, which defines the hyper parameter and other information required for training.

Our code use pytorch-lightning as the main framework and diffusers for loading pretrained stable-diffusion model. We mainly follow the implementation of stable-diffusion.

If you find our work useful in your research, please consider citing our paper:

@misc{ji2023improving,
      title={Improving Diffusion Models for Scene Text Editing with Dual Encoders}, 
      author={Jiabao Ji and Guanhua Zhang and Zhaowen Wang and Bairu Hou and Zhifei Zhang and Brian Price and Shiyu Chang},
      year={2023},
      eprint={2304.05568},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}