IOVST

Iterative Optimization Video Style Transfers by pytorch.

youtube : https://youtu.be/iDNP2KfThEg

blog(written in Japanese) : https://zenn.dev/kumamemo/articles/cca45744cc12de

Installation

Repository & Pretrained Weight

git clone https://github.com/hirokic5/IOVST.git
cd IOVST

Pretrained Models for RAFT & VGG

• RAFT (original repository)
• VGG

Dependencies

• PyTorch (>= 1.8.2)
• torchvision
• pytorch_cluster (==1.5.9)
• tqdm
• opencv-contrib

environment

Anaconda is recommended.

conda create -n iovst python=3.8
conda activate iovst

Libraries

pytorch

install torch / torchvision, reference to official website

pytorch-cluster

reference to official repository

In my environment, it didn't work with version 1.6.0 above...

Here is sample install command.

 pip install torch-cluster  https://data.pyg.org/whl/torch-1.8.0%2Bcu102/torch_cluster-1.5.9-cp38-cp38-win_amd64.whl

others

finally, install dependent libraries

pip install -r requirements.txt

Usage

Quick demo

python main.py --setting setting.py

After calculated, results is generated in results/

Tips

• More multipass iteration, better quality (and more time...).
• Less number of frames for stylized, less iteration is enough.
• If estimated optical flow is not good, quality of output is also not good...

Reference

Pytorch Brushstroke StyleTransfer implementation

RAFT official repository

Thesis : Artistic style transfer for videos