InstaGAN: Instance-aware Image-to-Image Translation
Warning: This repo contains a model which has potential ethical concerns. Remark that the task of jeans<->skirt was a bad application and should not be used in future research. See the twitter thread for the discussion.
PyTorch implementation of "InstaGAN: Instance-aware Image-to-Image Translation" (ICLR 2019). The implementation is based on the official CycleGAN code. Our major contributions are in ./models/insta_gan_model.py and ./models/networks.py.
Getting Started
Installation
- Clone this repository
git clone https://github.com/sangwoomo/instagan
- Install PyTorch 0.4+ and torchvision from http://pytorch.org and other dependencies (e.g., visdom and dominate). You can install all the dependencies by
pip install -r requirements.txt
-
For Conda users, you can use a script
./scripts/conda_deps.shto install PyTorch and other libraries. -
Acknowledgment: Installation scripts are from the official CycleGAN code.
Download base datasets
- Download clothing-co-parsing (CCP) dataset:
git clone https://github.com/bearpaw/clothing-co-parsing ./datasets/clothing-co-parsing
- Download multi-human parsing (MHP) dataset:
# Download "LV-MHP-v1" from the link and locate in ./datasets
- Download MS COCO dataset:
./datasets/download_coco.sh
Generate two-domain datasets
- Generate two-domain dataset for experiments:
python ./datasets/generate_ccp_dataset.py --save_root ./datasets/jeans2skirt_ccp --cat1 jeans --cat2 skirt
python ./datasets/generate_mhp_dataset.py --save_root ./datasets/pants2skirt_mhp --cat1 pants --cat2 skirt
python ./datasets/generate_coco_dataset.py --save_root ./datasets/shp2gir_coco --cat1 sheep --cat2 giraffe
- Note: Generated dataset contains images and corresponding masks, which are located in image folders (e.g., 'trainA') and mask folders (e.g., 'trainA_seg'), respectively. For each image (e.g., '0001.png'), corresponding masks for each instance (e.g., '0001_0.png', '0001_1.png', ...) are provided.
Run experiments
- Train a model:
python train.py --dataroot ./datasets/jeans2skirt_ccp --model insta_gan --name jeans2skirt_ccp_instagan --loadSizeH 330 --loadSizeW 220 --fineSizeH 300 --fineSizeW 200 --niter 400 --niter_decay 200
python train.py --dataroot ./datasets/pants2skirt_mhp --model insta_gan --name pants2skirt_mhp_instagan --loadSizeH 270 --loadSizeW 180 --fineSizeH 240 --fineSizeW 160
python train.py --dataroot ./datasets/shp2gir_coco --model insta_gan --name shp2gir_coco_instagan --loadSizeH 220 --loadSizeW 220 --fineSizeH 200 --fineSizeW 200
-
To view training results and loss plots, run
python -m visdom.serverand click the URL http://localhost:8097. To see more intermediate results, check out./checkpoints/experiment_name/web/index.html. -
For faster experiment, increase batch size and use more gpus:
python train.py --dataroot ./datasets/shp2gir_coco --model insta_gan --name shp2gir_coco_instagan --loadSizeH 220 --loadSizeW 220 --fineSizeH 200 --fineSizeW 200 --batch_size 4 --gpu_ids 0,1,2,3
- Test the model:
python test.py --dataroot ./datasets/jeans2skirt_ccp --model insta_gan --name jeans2skirt_ccp_instagan --loadSizeH 300 --loadSizeW 200 --fineSizeH 300 --fineSizeW 200
python test.py --dataroot ./datasets/pants2skirt_mhp --model insta_gan --name pants2skirt_mhp_instagan --loadSizeH 240 --loadSizeW 160 --fineSizeH 240 --fineSizeW 160 --ins_per 2 --ins_max 20
python test.py --dataroot ./datasets/shp2gir_coco --model insta_gan --name shp2gir_coco_instagan --loadSizeH 200 --loadSizeW 200 --fineSizeH 200 --fineSizeW 200 --ins_per 2 --ins_max 20
- The test results will be saved to a html file here:
./results/experiment_name/latest_test/index.html.
Apply a pre-trained model
-
You can download a pre-trained model (pants->skirt and/or sheep->giraffe) from the following Google drive link. Save the pretrained model in
./checkpoints/directory. -
We provide samples of two datasets (pants->skirt and sheep->giraffe) in this repository. To test the model:
python test.py --dataroot ./datasets/pants2skirt_mhp --model insta_gan --name pants2skirt_mhp_instagan --loadSizeH 240 --loadSizeW 160 --fineSizeH 240 --fineSizeW 160 --ins_per 2 --ins_max 20 --phase sample --epoch 200
python test.py --dataroot ./datasets/shp2gir_coco --model insta_gan --name shp2gir_coco_instagan --loadSizeH 200 --loadSizeW 200 --fineSizeH 200 --fineSizeW 200 --ins_per 2 --ins_max 20 --phase sample --epoch 200
Results
We provide some translation results of our model. See the link for more translation results.
1. Fashion dataset (pants->skirt)
2. COCO dataset (sheep->giraffe)
3. Results on Google-searched images (pants->skirt)
4. Results on YouTube-searched videos (pants->skirt)
Citation
If you use this code for your research, please cite our papers.
@inproceedings{
mo2019instagan,
title={InstaGAN: Instance-aware Image-to-Image Translation},
author={Sangwoo Mo and Minsu Cho and Jinwoo Shin},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=ryxwJhC9YX},
}
231 Jan 05, 2023
1 Dec 14, 2021
61 Jan 03, 2023
162 Nov 28, 2022
176 Jan 07, 2023
1.3k Jan 01, 2023
[email protected] Lab">
192 Jan 05, 2023
3 Sep 19, 2022
144 Dec 06, 2022
2.1k Dec 30, 2022
8 Dec 10, 2022
14 Dec 17, 2022
37 Oct 30, 2022
16 Jul 27, 2022
113 Dec 17, 2022
1 Mar 08, 2022
140 Jan 02, 2023
1 Dec 14, 2022
7 Nov 13, 2022
11 Dec 26, 2022