This is the official repository for the following paper:
Deep Image Harmonization with Learnable Augmentation [arXiv]
Li Niu, Junyan Cao, Wenyan Cong, Liqing Zhang
Accepted by ICCV 2023.
SycoNet can generate multiple plausible synthetic composite images based on a real image and a foreground mask, which is useful to construct pairs of synthetic composite images and real images for harmonization. We release the SycoNet inference code and model. The released model is first trained on iHarmony4 and then finetuned on ccHarmony, because ccHarmony can reflect the illumination variation more faithfully.
The released model could be used to generate high-quality synthetic composite images for real images to augment the small-scale training set.
Clone the repository:
git clone git@github.com:bcmi/SycoNet-Adaptive-Image-Harmonization.git
Install Anaconda and create a virtual environment:
conda create -n syconet python=3.6
conda activate syconet
Install PyTorch:
conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge
Install necessary packages:
pip install -r requirements.txt
Build Trilinear:
cd trilinear_cpp
sh setup.sh
Modify CUDA_HOME as your own path in setup.sh. You can refer to this repository for more solutions.
Download SycoNet model pretrained_net_Er.pth and 3D LUTs pretrained_net_LUTs.pth from Baidu Cloud (access code:o4rt) or GoogleDrive. Put them in the folder checkpoints\syco.
Modify real and mask in demo_test.sh as your own real image path and foreground mask path respectively. Modify augment_num as your expected number of generated composite images per pair of real image and foreground mask. Then, run the following command:
sh demo_test_single.sh
Our SycoNet could generate composite images for the input real image and foreground mask in the folder results\syco\test_pretrained.
Download iHarmony4 and modify dataset_root, dataset_name in demo_test_iHarmony4.sh as your own dataset path. Then, run the following command:
sh demo_test_iHarmony4.sh
Our SycoNet could generate composite images for the input real images and foreground masks in the specified dataset in the folder results\syco\test_pretrained.