Official repository of the paper (AI for Content Creation (AICC) Workshop, CVPR 2022) "Image2Gif: Generating Continuous Realistic Animations with Warping NODEs", Warping Neural ODE.

If you like our work, please give us a star. If you use our work in your research projects, please cite our paper as

@article{nazarovs2022image2gif,
  title={Image2Gif: Generating Continuous Realistic Animations with Warping NODEs},
  author={Nazarovs, Jurijs and Huang, Zhichun},
  journal={arXiv preprint arXiv:2205.04519},
  year={2022}
}

Installation

Packages

The requirements.txt was generated with pipreqs --force ./ > requirements.txt. It contains description of packages used for this project. You can install required packages as pip install -r requirements.txt. Note: requirements.txt was not generated as a copy-slice from conda environment.

Data

The demo data is provided in ./datasets/ folder. You need to extract data as tar -xzvf RotatingMnist.tar.gz and tar -xzvf smile.tar.gz.

Running

There are 2 files, which provide all options to the code.

1. data_config.py describes all information about data. If you would like to add your own set, you can do it by adding extra elif, with corresponding name of your dataset and 4 parameters: source path, target path, image size and number of channels. Note that some images have 4 channels, which include RGB and alpha. Currently, network works with images of size 32 and 64. If you provide images with higher resolution, they will be resized to 64x64 pixels, in utils.py function get_data.

2. config.py describes all parameters used for the network. They can be either changed in that file or provided as argument in the terminal. Examples below.

Visualization of experiments

Warping MNIST

To train the model on warping from 3 to 4, you can run the folloowing code:

python3 main.py --n_epochs 5000 --data mnist_warp --experimentID mnist_warp --device 0  --batch_size 64  --n_epochs_start_viz 50000 --gen_loss_weight 1 --disc_loss_weight 1 --gan_type lsgan --ode_solver euler  --freq_gen_update 1 --decay_every 100 --lr_gen 0.0001 --lr_disc 0.0001 --ode_vf init_img_y --ode_norm none --plots_path plots/ --rec_loss_weight 0.01 --freq_rec_update 1 --last_warp --jac_loss_weight_forw 1 --jac_loss_weight_back 1 --rec_weight_method default --outgrid_loss_weight_forw 1 --outgrid_loss_weight_back 1 --crit_iter 10 --gplambda 10 --disc_optim adam --unet_add_input --normalize_method scale --ode_step_size 0.05 --channels 1

Once training is done, you can add to the previous command the following arguments: --load --test_only --plot_all. It will load the last model, and produce images of warping in the following directory: plots/mnnist_warp/vf_seq/.

Then you can use make_gif.sh script to generate gif as (for example): ./makegif.sh plots_gif/mnist.gif plots/mnist_warp/vf_seq/0_0/ 2 100.

Smiling face

Similar, to generate smiling faces you can run the following code:

python3 main.py --n_epochs 5000 --data face_warp_smile --experimentID face_warp_smile --device 0  --batch_size 16  --n_epochs_start_viz 50000 --gen_loss_weight 1 --disc_loss_weight 1 --gan_type None --ode_solver euler  --freq_gen_update 1 --decay_every 100 --lr_gen 0.0001 --lr_disc 0.0001 --ode_vf init_img_y --ode_norm none --plots_path plots/ --rec_loss_weight 0.01 --freq_rec_update 1 --last_warp --jac_loss_weight_forw 1 --jac_loss_weight_back 1 --rec_weight_method default --outgrid_loss_weight_forw 1 --outgrid_loss_weight_back 1 --crit_iter 10 --gplambda 10 --disc_optim adam --unet_add_input --normalize_method scale --ode_step_size 0.05 --channels 4 

Once training is done, you can add to the previous command the following arguments: --load --test_only --plot_all . It will load the last model, and produce images of warping in the following directory: plots/face_warp_smile/vf_seq/.

Then you can use make_gif.sh script to generate gif as (for example): ./makegif.sh plots_gif/smile.gif plots/face_warp_smile/vf_seq/0_0/ 2 100.