This repository implements the paper "Text-Guided Neural Image Inpainting" by Lisai Zhang, Qingcai Chen, Baotian Hu and Shuoran Jiang. Given one masked image, the proposed TDANet generates diverse plausible results according to guidance text.

Inpainting example


Manipulation Extension example


Getting started


This code was tested with Pytoch 1.2.0, CUDA 10.1, Python 3.6 and Ubuntu 16.04 with a 2080Ti GPU

pip install visdom dominate
  • Clone this repo (we suggest to only clone the depth 1 version):
git clone --depth 1
cd tdanet
  • Download the dataset and pre-processed files as in following steps.


  • CUB_200: dataset from Caltech-UCSD Birds 200.
  • COCO: object detection 2014 datset from MS COCO.
  • pre-processed datafiles: train/test split, caption-image mapping, image sampling
    and pre-trained DAMSM from GoogleDrive and extarct them
    to dataset/ directory as specified in config.bird.yml/config.coco.yml.

Training Demo

python --name tda_bird  --gpu_ids 0 --model tdanet --mask_type 0 1 2 3 --img_file ./datasets/CUB_200_2011/train.flist --mask_file ./datasets/CUB_200_2011/train_mask.flist --text_config config.bird.yml
  • Important: Add --mask_type in options/ for different training masks. --mask_file path is needed for object mask, use train_mask.flist for CUB and image_mask_coco_all.json for COCO. --text_config refer to the yml configuration file for text setup, --img_file is the image file dir or file list.
  • To view training results and loss plots, run python -m visdom.server and copy the URL http://localhost:8097.
  • Training models will be saved under the ./checkpoints folder.
  • More training options can be found in ./options folder.
  • Suggestion: use mask type 0 1 2 3 for CUB dataset and 0 1 2 4 for COCO dataset. Train more than 2000 epochs for CUB and 200 epochs for COCO.

Evaluation Demo


python --name tda_bird  --img_file datasets/CUB_200_2011/test.flist --results_dir results/tda_bird  --mask_file datasets/CUB_200_2011/test_mask.flist --mask_type 3 --no_shuffle --gpu_ids 0 --nsampling 1 --no_variance


  • Remember to add the --no_variance option to get better performance.
  • For COCO object mask, use image_mask_coco_all.json as the mask file..

A eval_tda_bird.flist will be generated after the test. Then in the evaluation, this file is used as the ground truth file list:

python --batch_test 60 --ground_truth_path eval_tda_bird.flist --save_path results/tda_bird
  • Add --ground_truth_path to the dir of ground truth image path or list. --save_path as the result dir.

Pretrained Models

Download the pre-trained models bird inpainting or coco inpainting and put them undercheckpoints/ directory.


  • Install the PyQt5 for GUI operation
pip install PyQt5

The GUI could now only avaliable in debug mode, please refer to this issues for detailed instructions. The author is not good at solving PyQt5 problems, wellcome contrbutions.


  • Debug the GUI application
  • Further improvement on COCO quality.


This software is for educational and academic research purpose only. If you wish to obtain a commercial royalty bearing license to
this software, please contact us at [email protected].


We would like to thanks Zheng et al. for providing their source code. This project is fit from their greate Pluralistic Image Completion Project.


If you use this code for your research, please cite our paper.

author = {Zhang, Lisai and Chen, Qingcai and Hu, Baotian and Jiang, Shuoran},
title = {Text-Guided Neural Image Inpainting},
year = {2020},
booktitle = {Proceedings of the 28th ACM International Conference on Multimedia},
pages = {1302–1310},
location = {Seattle, WA, USA},