This is the pytorch implementation of Aesthetic-Aware Image Style Transfer [Hu et al., MM2020].
Our work propose a novel problem called Aesthetic-Aware Image Style Transfer(AAST) which aims to control the texture and color of an image independently during style transfer and thus generate results with more diverse aesthetic effects.
Our code framework refers to the framework of pytorch-AdaIN.
- python 3.7+
- pytorch 1.4+
- torchvision 0.5+
- Pillow
(Optional)
-
CUDA 10.0
-
tqdm
-
TensorboardX
Download vgg_normalized.pth
(required for training) and net_final.pth
(for testing only) and put them under models/
vgg_normalized.pth
:
net_final.pth
:
Use --content_dir
, --texture_dir
and --color_dir
to specify directories that save content images, texture reference images and color reference images. The model will iterate over all combinations between content, texture and color.
Use --test_opt
to specify the type of test you want to conduct:
TC
: Transfer texture and color together.ST
: Traditional style transfer.T
: Texture only transfer.C
: Color only transfer.INT
: Parameter interpolation for texture and color. For this type of test, you need to specify the interpolation num by--int_num
python main.py --mode test --test_opt <TEST_OPTION> --content_dir <CONTENT_DIR> --texture_dir <TEXTURE_DIR> --color_dir <COLOR_DIR>
For more detailed configurations, please refer to --help
option.
Use --content_dir
, --texture_dir
and --color_dir
to specify directories that save content images, texture reference images and color reference images.
In each iteration, the model will randomly sample a batch of content-texture-color pair for training. The training will stop when it reaches the maximum iteration num (specified by --max_iter
). Usually the training will not iterate over the whole dataset as the num of combinations of content, texture and color is really large.
python main.py --mode train --content_dir <CONTENT_DIR> --texture_dir <TEXTURE_DIR> --color_dir <COLOR_DIR>
For more detailed configurations, please refer to --help
option.
If you have any questions or suggestions about this paper, feel free to contact me (z8hu@ucsd.edu).