StyleGAN-V: A Continuous Video Generator with the Price, Image Quality and Perks of StyleGAN2

[CVPR 2022] Official pytorch implementation

[Project website] [Paper] [Casual GAN papers summary]

Code release TODO:

• Installation guide
• Training code
• Data preprocessing scripts
• CLIP editing scripts (50% done)
• Jupyter notebook demos
• Pre-trained checkpoints

Installation

To install and activate the environment, run the following command:

conda env create -f environment.yaml -p env
conda activate ./env

For clip editing, you will need to install StyleCLIP and clip. This repo is built on top of INR-GAN, so make sure that it runs on your system.

If you have Ampere GPUs (A6000, A100 or RTX-3090), then use environment-ampere.yaml instead because it is based CUDA 11 and newer pytorch versions.

System requirements

Our codebase uses the same system requirements as StyleGAN2-ADA: see them here. We trained all the 256x256 models on 4 V100s with 32 GB each for ~2 days. It is very similar in training time to StyleGAN2-ADA (even a bit faster).

Training

Dataset structure

The dataset should be either a .zip archive (the default setting) or a directory structured as:

dataset/
    video1/
        - frame1.jpg
        - frame2.jpg
        - ...
    video2/
        - frame1.jpg
        - frame2.jpg
        - ...
    ...

We use such frame-wise structure because it makes loading faster for sparse training.

By default, we assume that the data is packed into a .zip archive since such representation is useful to avoid additional overhead when copying data between machines on a cluster. You can also train from a directory: for this, just remove the .zip suffix from the dataset.path property in configs/dataset/base.yaml.

If you want to train on a custom dataset, then create a config for it here configs/dataset/my_dataset_config_name.yaml (see configs/dataset/ffs.yaml as an example). The fps parameter is needed for visualizations purposes only, videos typically have the value of 25 or 30 FPS.

Training StyleGAN-V

To train on FaceForensics 256x256, run:

python src/infra/launch.py hydra.run.dir=. exp_suffix=my_experiment_name env=local dataset=ffs dataset.resolution=256 num_gpus=4

To train on SkyTimelapse 256x256, run:

python src/infra/launch.py hydra.run.dir=. exp_suffix=my_experiment_name env=local dataset=sky_timelapse dataset.resolution=256 num_gpus=4 model.generator.time_enc.min_period_len=256

For SkyTimelapse 256x256, we increased the period length for the motion time encoder since the motions in this dataset are much slower/smoother, than in FaceForensics. In practice, this parameter (and its accompanying model.generator.motion.motion_z_distance) influences the motion quality (but not the image quality!) the most.

If you do not want hydra to create some log directories (typically, you don't), add the following arguments: hydra.output_subdir=null hydra/job_logging=disabled hydra/hydra_logging=disabled.

In case slurm is installed on your system, you can submit the slurm job with the above training by adding slurm=true parameter. Sbatch arguments are specified in configs/infra.yaml, you can update them with your required ones. Also note that you can create your own environment in configs/env.

On older GPUs (non V100 and newer), custom CUDA kernels (bias_act and upfirdn2n) might fail to compile. The following two lines can help:

export TORCH_CUDA_ARCH_LIST="7.0"
export TORCH_EXTENSIONS_DIR=/tmp/torch_extensions

Resume training

If you shut down your experiment at some point and would love to fully recover training (i.e., with the optimizer parameters, logging, etc.), the add training.resume=latest argument to your launch script, e.g.:

python src/infra/launch.py hydra.run.dir=. exp_suffix=my_experiment_name env=local dataset=ffs dataset.resolution=256 num_gpus=4 training.resume=latest

It will locate the given experiment directory (note that the git hash and the exp_suffix must be the same) and resume the training from it.

Inference

To sample from the model, launch the following command:

python src/scripts/generate.py --network_pkl /path/to/network-snapshot.pkl --num_videos 25 --as_grids true --save_as_mp4 true --fps 25 --video_len 128 --batch_size 25 --outdir /path/to/output/dir --truncation_psi 0.9

This will sample 25 videos of 25 FPS as a 5x5 grid with the truncation factor of 0.9. Each video consists of 128 frames. Adjust the corresponding arguments to change the settings.

Alternatively, instead of specifying --network_pkl, you can specify --networks_dir, which should lead to a directory containing the checkpoints and the metric-fvd2048_16f.json metrics json file (it is generated automatically during training). It will then select the best checkpoint based on the metrics, which so not to search for the best checkpoint of an experiment manually.

To sample content/motion decomposition grids, use --moco_decomposition 1 by running the following command:

python src/scripts/generate.py --networks_dir PATH_TO_EXPERIMENT/output --num_videos 25 --as_grids true --save_as_mp4 true --fps 25 --video_len 128 --batch_size 25 --outdir tmp --truncation_psi 0.8 --moco_decomposition 1

Training MoCoGAN + SG2 backbone

To train the MoCoGAN+SG2 model, just use the mocogan.yaml model config with the uniform sampling:

python src/infra/launch.py hydra.run.dir=. +exp_suffix=my_experiment env=local dataset=sky_timelapse dataset.resolution=256 num_gpus=4 model=mocogan sampling=uniform sampling.max_dist_between_frames=1

Training other baselines

To train other baselines, used in the paper, we used their original implementations:

• MoCoGAN
• MoCoGAN-HD
• DIGAN
• VideoGPT

Data

Datasets can be downloaded here:

• SkyTimelapse: https://github.com/weixiong-ur/mdgan
• UCF: https://www.crcv.ucf.edu/data/UCF101.php
• FaceForensics: https://github.com/ondyari/FaceForensics
• RainbowJelly: https://www.youtube.com/watch?v=P8Bit37hlsQ
• MEAD: https://wywu.github.io/projects/MEAD/MEAD.html

We resize all the datasets to the 256x256 resolution (except for MEAD, which we resize to 1024x1024). FFS was preprocessed with src/scripts/preprocess_ffs.py to extract face crops. For MEAD, we used only the front views.

For RainbowJelly, download the youtube video, save it as rainbow_jelly.mp4 and convert into the dataset by running:

python src/scripts/convert_video_to_dataset.py -s /path/to/rainbow_jelly.mp4 -t /path/to/desired/directory --target_size 256 -sf 150 -cs 512

Evaluation

In this repo, we re-implemented two popular evaluation measures for video generation:

• Frechet Video Distance. For this, we re-implemented perfectly (up to numerical precision) the original Tensorflow version of the I3D model trained on Kinetics-400 and converted it to TorchScript. This is a precise implementation of the official one and we set up this comparison repo to demonstrate this.
• Inception Score (used only for UCF101). For this, we re-implemented perfectly (up to numerical precision) the original Chainer version of the UCF101-finetuned C3D model in Pytorch and converted it to TorchScript.

In practice, we found that neither Frechet Video Distance nor Inception Score work well reliably for catching motion artifacts. This creates the need for better evaluation measures.

Advantages of our metrics implementation compared to the original ones:

• It is much faster due to TorchScript and parallelization across several GPUs
• It can be launched both on top of both a generator checkpoint and off-the-shelf samples
• It is implemented in a very recent Pytorch version (v1.9.0) instead of deprecated TensorFlow 1.14 or Chainer 6.0
• It is directly incorporated into training to track progress online without the need to launch the evaluation separately
• For FVD, our implementation is complete, while the original one provides evaluation for a single batch of already processed videos only
• Our FVD implementation supports different subsampling strategies and a variable number of frames in a video.

To compute FVD between two datasets, run the following command:

python src/scripts/calc_metrics_for_dataset.py --real_data_path /path/to/dataset_a.zip --fake_data_path /path/to/dataset_b.zip --mirror 1 --gpus 4 --resolution 256 --metrics fvd2048_16f,fvd2048_128f,fvd2048_128f_subsample8f,fid50k_full --verbose 0 --use_cache 0

To compute FVD for a trained model, run src/scripts/calc_metrics.py instead.

Both datasets should be in the format specified above. They can be either zip archives or normal directories. This will compute several metrics:

• fid50k_full - Frechet Inception Distance
• fvd2048_16f — Frechet Video Distance with 16 frames
• fvd2048_128f - Frechet Video Distance with 128 frames
• fvd2048_128f_subsample8f — Frechet Video Distance with 16 frames, but sampled with a 8-frames interval

Note. If you face any trouble running the above evaluation scripts — please do not hesitate contacting us!

Projection and CLIP editing

This section is still under construction. We will update it shortly.

Those two files provide projection and editing scripts:

• src/scripts/project.py
• src/scripts/clip_edit.py

Infrastructure and visualization

You will find some useful scripts for data processing and visualization in src/scripts

Troubleshooting

Make sure that INR-GAN and StyleGAN2-ADA are runnable on your system. We do not use any additional CUDA kernels or any exotic dependencies.

If this didn't help, than it's likely there is some dependency version mismatch. Check the versions of your installed dependencies with:

pip freeze

and compare them with the ones specified in environment.yaml/environment-ampere.yaml.

If this didn't help — open an issue, it's likely that the problem is on our side.

License

This repo is built on top of INR-GAN, which is likely to be restricted by the NVidia license since it's built on top of StyleGAN2-ADA. If that's the case, then this repo is also restricted by it.

Bibtex

@misc{stylegan_v,
    title={StyleGAN-V: A Continuous Video Generator with the Price, Image Quality and Perks of StyleGAN2},
    author={Ivan Skorokhodov and Sergey Tulyakov and Mohamed Elhoseiny},
    journal={arXiv preprint arXiv:2112.14683},
    year={2021}
}

@inproceedings{digan,
    title={Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks},
    author={Sihyun Yu and Jihoon Tack and Sangwoo Mo and Hyunsu Kim and Junho Kim and Jung-Woo Ha and Jinwoo Shin},
    booktitle={International Conference on Learning Representations},
    year={2022},
    url={https://openreview.net/forum?id=Czsdv-S4-w9}
}