Text2Poster-ICASSP-22

The inference code of the ICASSP-2022 paper "Text2Poster: Laying Out Stylized Texts on Retrieved Images".

framework

Paper Link: https://arxiv.org/abs/2301.02363

Star History

If you like this project, please give it a star! It would be a great encouragement for me and help me to continue improving it.

Quick Start from API

Just run the following code to quick start:

import time, json, requests
timestamp = time.strftime('%Y%m%d%H%M%S',time.localtime(time.time())) 

input_text_elements = {
    "sentences": [
        ["CHILDREN'S DAY", 90], # [text, font_size]
        ["Children are The Future of Nation", 50] # [text, font_size]
    ],
    "background_query": "Children's Day!" # sentence used to retrieve background images.
}

input_text_elements = json.dumps(input_text_elements)
api_url = "http://bl.mmd.ac.cn:8889/text2poster"
response = requests.get(api_url, params = {"input_text_elements": input_text_elements})
if response.status_code == 200:
    f = open("poster-{}.jpg".format(timestamp), "wb")
    f.write(response.content)
    f.close()
    print("Save poster to:", "poster-{}.jpg".format(timestamp))
else:
    print(response.text)

News

[2023.1.24] Update "http://1.13.255.9" to "http://bl.mmd.ac.cn".
[2023.1.17] We provide an API for Text2Poster, you can quickly start our Text2Poster without any resource download.
[2023.1.16] We add a machine translation API to translate all the input texts into Chinese. So that the BriVL model could deal with all languages. Now you can try to retrieve background images in any language!!
[2023.1.15] We provide an unsplash image download script in ./background_retriever/unsplash_image_downloader.py, you can use this script to get those background image files based on the retrieved image urls.
[2023.1.14] We provide our image retrieval source code and data in ./background_retriever for the convenience of people who are not in mainland China to use our Text2Poster. You can retrieve the background image locally, which requires about 3GB of GPU memory.
[2023.01.10] We update the background image retrieval website to http://1.13.255.9:8889. The original website buling.wudaoai.cn has been retired in 2023.01.09.

Generated Posters:

poster poster

More Examples

input text elements 1

55, 40 and 30 are the font size.

{
    "sentences": [
        ["冬日初雪舞会", 55],
        ["雪花飞舞，像音乐和歌声围绕", 40],
        ["与朋友相聚，享受欢乐时光，我们不见不散", 30]
    ],
    "background_query": "冬日初雪舞会"
}

output posters

poster poster

input text elements 2

80 and 55 are the font size.

{
    "sentences": [
        ["ICASSP 2022", 80],
        ["May 22 - 27, 2022, Singapore", 55]
    ],
    "background_query": "Singapore"
}

output posters

poster poster

input text elements 3

90 and 50 are the font size.

{
    "sentences": [
        ["桜が咲く", 90],
        ["出会いは素晴らしい春に", 50]
    ],
    "background_query": "春の美しい桜"
}

output posters

poster poster

Start from Source Code

Install

We recommend you use anaconda to run our Text2Poster. Run the following command to install the dependent libraries:

bash install_package.sh

you also can install the dependent libraries manually:

# using the tsinghua mirror to speed up the install.
conda install pytorch=1.10.0 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/
conda install torchvision=0.11.0 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/
pip install opencv_contrib_python
pip install transformers==3.2.0
pip install argparse
pip install freetype-py
pip install requests
pip install jsonlines
pip install tqdm
pip install pyyaml
pip install easydict
pip install timm

Download

We provide the following resource to start Text2Poster:

Weights of layout refine model: ./checkpoint/0.20484_Cascading_128_uniform_big.pth;
Weights of layout distribution prediction model: ./checkpoint/27.80619_distribCNN_BigPosition_epoch_76_scale_20.pth;

[Not required] Although we provide an API for background image retrieval, if you want to retrieve background images from the source code, you need to download the following resources:

Weights of text encoder of BriVL: brivl-textencoder-weights.pth -> ./background_retriever/weights/;
Unsplash images features (extracted by BriVL): wenlan_unsplash_feats.npy -> ./background_retriever/background_feats/;
URL of background images: ./background_retriever/background_feats/unsplash_image_url.jsonl.

Running

We provide two example, Run the following command to run our Text2Poster:

bash run.sh

Some parameters:

input_text_file: The input text elements, it contains: 1). sentences (phase) and their font size, 2). query used to retrieve background images.
output_folder: The folder to save the output posters and some process figures.
background_folder: The folder to save local background images, If images are not saved locally, they will be downloaded from remote.
top_n: Arrange the text elements on the top N retrieved images.
save_process: Save the process figure (etc. saliency map) or not.

We also provide the following examples:

background image retrieval （from API）

python background_retrieval.py

background image retrieval （from source code）

cd background_retriever
python main.py

Layout distribution prediction

python layout_distribution_predict.py

Layout refinement

python layout_refine.py

Download images from Unsplash

python ./background_retriever/unsplash_image_downloader.py

Some Output During Process

we also output some intermediate processing files in ./example/outputs:

-SdD0KbD7N0 saliency_map_with-smooth

Right image: The original background image.
Left image: Saliency map (blue) with smooth region map (red).

layout_distribution saliency_map_with-smooth

Right image: The prediction of layout distribution map.
Left image: Saliency map (blue) with predicted layout distribution map (red).

initial_layout refined_layout

Right image: Initial layout map.
Left image: Refined layout map.

Blue region: The saliency map;

Green region: The predicted layout distribution map;

Red region: the predicted layout map.

Tips

Something about our background image retrieval

Our background image retrieval is implemented by a Chinese pre-trained text-image retrieval model BriVL.
~~You can extract text and image embedding by the API of BriVL at here.~~ The API server has been retired in 2023.01.09.
A text-image retrieval application is provided at http://1.13.255.9:8889. The core code of this application at https://github.com/chuhaojin/BriVL-BUA-applications.

Requirements

python==3.7
pytorch=1.10.0
torchvision=0.11.0
transformers==3.2.0
freetype-py
opencv_contrib_python
requests
jsonlines
tqdm
argparse
pyyaml
easydict
timm

Citation

If you find this paper and repo useful, please cite us in your work:

@inproceedings{DBLP:conf/icassp/JinXSL22,
  author    = {Chuhao Jin and
               Hongteng Xu and
               Ruihua Song and
               Zhiwu Lu},
  title     = {Text2Poster: Laying Out Stylized Texts on Retrieved Images},
  booktitle = {{IEEE} International Conference on Acoustics, Speech and Signal Processing,
               {ICASSP} 2022, Virtual and Singapore, 23-27 May 2022},
  pages     = {4823--4827},
  publisher = {{IEEE}},
  year      = {2022},
  url       = {https://doi.org/10.1109/ICASSP43922.2022.9747465},
  doi       = {10.1109/ICASSP43922.2022.9747465},
  timestamp = {Tue, 07 Jun 2022 17:34:56 +0200},
  biburl    = {https://dblp.org/rec/conf/icassp/JinXSL22.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Contact

My Email is: jinchuhao@ruc.edu.cn