PyTorch implementation for paper "SAC-GAN: Structure-Aware Image Composition", Hang Zhou, Rui Ma, Ling-Xiao Zhang, Lin Gao, Ali Mahdavi-Amiri, Hao (Richard) Zhang.
Our code has been tested with Python 3.6, Pytorch 1.6.0, CUDA 10.1 and cuDNN 7.0 on Ubuntu 18.04.
Install required dependencies, run
conda env create -f environment.yml
Our network is trained individually on Cityscapes (for vehicle/pedestrian/street light&sign composition), 3D-FUTURE & 3D-FRONT rendered images (for chair composition), and CelebAHQ (for glasses composition).
Pretrained model of vehicle composition: GoogleDrive
Train SAC-GAN on Cityscapes:
python main.py \
--phase train \
--dataset_name 'cityscapes' \
--dataset YOUR_DATA_DIR \
--result_dir YOUR_RESULTS_DIR \
--epoch 10 \
--class_num 19 \
--layout_flag True \
--save_freq 500 \
--rec_weight 100 \
--latent_rec_weight 0.05 \
--affine_weight 1 \
--layout_weight 1
Test
python main.py \
--phase test \
--dataset_name 'cityscapes' \
--dataset YOUR_DATA_DIR \
--result_dir YOUR_RESULTS_DIR \
--class_num 19 \
--layout_flag False
This project is licensed under the terms of the MIT license.