Unofficial DLP-GAN Implementation

Unofficial PyTorch implementation of DLP-GAN: learning to draw modern Chinese landscape photos with generative adversarial network

For the official implementation, please refer to the original paper.

Installation

# Clone this repo
git clone https://github.com/injea/AncientPainitng2NaturalImage.git
cd AncientPainitng2NaturalImage

# Install dependencies
pip install -r requirements.txt

Dataset

Since the original DLP-GAN paper doesn't provide the exact dataset used, we use modern landscape photos from LHQ-1024 dataset for training. This is an unofficial implementation and may differ from the original paper's implementation.

Manual Download

• Download datasets (e.g. CBP, CFP) : https://drive.google.com/open?id=1ilqfMC3A9Kt6CaoZZCT9tI-wWRl1kLFB
• Chinese Landscape Painting Dataset : https://github.com/alicex2020/Chinese-Landscape-Painting-Dataset
• LHQ-1024 (Modern landscape photos) : https://www.kaggle.com/datasets/dimensi0n/lhq-1024

Automatic Download (Recommended)

You can automatically download all datasets using our built-in downloader:

# Download all datasets at once
python util/get_data.py

# Or download specific datasets
python -c "
from util.get_data import GetData
gd = GetData()

# Download DLP-GAN dataset (Chinese landscape paintings)
gd.get_dlp_gan_dataset('./datasets/dlp_gan_dataset')

# Download DSTN datasets (Bird, Flower, Landscape)
gd.get_dstn_dataset('./datasets/dstn_dataset')

# Download LHQ-1024 dataset (Modern landscape photos)
gd.get_lhq_1024_dataset('./datasets/lhq_1024')
"

The datasets will be automatically organized in the ./datasets/ directory with the proper structure for training.

Citation

@article{gui2023dlp,
  title        = {DLP-GAN: Learning to Draw Modern Chinese Landscape Photos with Generative Adversarial Network},
  author       = {Gui, Xiangquan and Zhang, Binxuan and Li, Li and Yang, Yi},
  journal      = {Neural Computing and Applications},
  pages        = {1--18},
  year         = {2023},
  publisher    = {Springer},
  doi          = {10.1007/s00521-023-09345-8}
}

Acknowledgments

Code is inspired by AncientPainitng2NaturalImage.

Notice

⚠️ Important Implementation Fix: The original AncientPainitng2NaturalImage repository has a critical bug in VGG input processing. It subtracts ImageNet mean values [103.939, 116.779, 123.680] from images that are already normalized to the -1~1 scale, resulting in input images with a problematic -130~-100 scale range being fed into the VGG network.

In this implementation, I have replaced the VGG model with PyTorch's pretrained VGG and fixed this scaling issue. Please be aware of this difference when comparing results. You can validate this VGG preprocessing fix using the provided imagenet_vgg_evaluation.py script, which compares classification accuracy between the standard ImageNet normalization (correct approach) and the pixel-wise mean subtraction (original buggy approach). The validation dataset used is the ImageNet Train Subset (100k images) from Kaggle: https://www.kaggle.com/datasets/tusonggao/imagenet-train-subset-100k