Stitch it in Time: GAN-Based Facial Editing of Real Videos
Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, Daniel Cohen-Or
Abstract:
The ability of Generative Adversarial Networks to encode rich semantics within their latent space has been widely adopted for facial image editing. However, replicating their success with videos has proven challenging. Sets of high-quality facial videos are lacking, and working with videos introduces a fundamental barrier to overcome - temporal coherency. We propose that this barrier is largely artificial. The source video is already temporally coherent, and deviations from this state arise in part due to careless treatment of individual components in the editing pipeline. We leverage the natural alignment of StyleGAN and the tendency of neural networks to learn low frequency functions, and demonstrate that they provide a strongly consistent prior. We draw on these insights and propose a framework for semantic editing of faces in videos, demonstrating significant improvements over the current state-of-the-art. Our method produces meaningful face manipulations, maintains a higher degree of temporal consistency, and can be applied to challenging, high quality, talking head videos which current methods struggle with.
- Pytorch(tested with 1.10, should work with 1.8/1.9 as well) + torchvision, Follow https://pytorch.org/ for pytorch installation instructions.
- CUDA toolkit 11.0 or later. Make sure you have the requirements listed here: https://github.com/NVlabs/stylegan2-ada-pytorch#requirements
- For the rest of the requirements, use:
pip install -r requirements.txt - To perform StyleCLIP edits, install clip with:
pip install git+https://github.com/openai/CLIP.git
In order to use this project you need to download pretrained models from the following Link.
Unzip it inside the project's main directory.
You can use the download_models.sh script (requires installing gdown with pip install gdown)
Alternatively, you can unzip the models to a location of your choice and update configs/path_config.py accordingly.
Our code expects videos in the form of a directory with individual frame images. To produce such a directory from an existing video, we recommend using ffmpeg:
ffmpeg -i "video.mp4" "video_frames/out%04d.png"
The videos used to produce our results can be downloaded from the following Link.
To invert a video run:
python train.py --input_folder /path/to/images_dir \
--output_folder /path/to/experiment_dir \
--run_name RUN_NAME \
--num_pti_steps NUM_STEPS
This includes aligning, cropping, e4e encoding and PTI
For example:
python train.py --input_folder /data/obama \
--output_folder training_results/obama \
--run_name obama \
--num_pti_steps 80
Weights and biases logging is disabled by default. to enable, add --use_wandb
To run edits without stitching tuning:
python edit_video.py --input_folder /path/to/images_dir \
--output_folder /path/to/experiment_dir \
--run_name RUN_NAME \
--edit_name EDIT_NAME \
--edit_range EDIT_RANGE
edit_range determines the strength of the edits applied.
It should be in the format RANGE_START RANGE_END RANGE_STEPS.
for example, if we use --edit_range 1 5 2,
we will apply edits with strength 1, 3 and 5.
For young Obama use:
python edit_video.py --input_folder /data/obama \
--output_folder edits/obama/ \
--run_name obama \
--edit_name age \
--edit_range -8 -8 1 \
To run edits with stitching tuning:
python edit_video_stitching_tuning.py --input_folder /path/to/images_dir \
--output_folder /path/to/experiment_dir \
--run_name RUN_NAME \
--edit_name EDIT_NAME \
--edit_range EDIT_RANGE \
--outer_mask_dilation MASK_DILATION
We support early breaking the stitching tuning process, when the loss reaches a specified threshold.
This enables us to perform more iterations for difficult frames while maintaining a reasonable running time.
To use this feature, add --border_loss_threshold THRESHOLD to the command(Shown in the Jim and Kamala Harris examples below).
For videos with a simple background to reconstruct (e.g Obama, Jim, Emma Watson, Kamala Harris), we use THRESHOLD=0.005.
For videos where a more exact reconstruction of the background is required (e.g Michael Scott), we use THRESHOLD=0.002.
Early breaking is disabled by default.
For young Obama use:
python edit_video_stitching_tuning.py --input_folder /data/obama \
--output_folder edits/obama/ \
--run_name obama \
--edit_name age \
--edit_range -8 -8 1 \
--outer_mask_dilation 50
For gender editing on Obama use:
python edit_video_stitching_tuning.py --input_folder /data/obama \
--output_folder edits/obama/ \
--run_name obama \
--edit_name gender \
--edit_range -6 -6 1 \
--outer_mask_dilation 50
For young Emma Watson use:
python edit_video_stitching_tuning.py --input_folder /data/emma_watson \
--output_folder edits/emma_watson/ \
--run_name emma_watson \
--edit_name age \
--edit_range -8 -8 1 \
--outer_mask_dilation 50
For smile removal on Emma Watson use:
python edit_video_stitching_tuning.py --input_folder /data/emma_watson \
--output_folder edits/emma_watson/ \
--run_name emma_watson \
--edit_name smile \
--edit_range -3 -3 1 \
--outer_mask_dilation 50
For Emma Watson lipstick editing use: (done with styleclip global direction)
python edit_video_stitching_tuning.py --input_folder /data/emma_watson \
--output_folder edits/emma_watson/ \
--run_name emma_watson \
--edit_type styleclip_global \
--edit_name lipstick \
--neutral_class "Face" \
--target_class "Face with lipstick" \
--beta 0.2 \
--edit_range 10 10 1 \
--outer_mask_dilation 50
For Old + Young Jim use (with early breaking):
python edit_video_stitching_tuning.py --input_folder datasets/jim/ \
--output_folder edits/jim \
--run_name jim \
--edit_name age \
--edit_range -8 8 2 \
--outer_mask_dilation 50 \
--border_loss_threshold 0.005
For smiling Kamala Harris:
python edit_video_stitching_tuning.py \
--input_folder datasets/kamala/ \
--output_folder edits/kamala \
--run_name kamala \
--edit_name smile \
--edit_range 2 2 1 \
--outer_mask_dilation 50 \
--border_loss_threshold 0.005
With stitching tuning:
out.mp4
Without stitching tuning:
out.mp4
Gender editing:
out.mp4
Young Emma Watson:
out.mp4
Emma Watson with lipstick:
out.mp4
Emma Watson smile removal:
out.mp4
Old Jim:
out.mp4
Young Jim:
out.mp4
Smiling Kamala Harris:
out.mp4
For editing out of domain videos, Some different parameters are required while training.
First, dlib's face detector doesn't detect all animated faces, so we use a different face detector provided by the face_alignment package.
Second, we reduce the smoothing of the alignment parameters with --center_sigma 0.0
Third, OOD videos require more training steps, as they are more difficult to invert.
To train, we use:
python train.py --input_folder datasets/ood_spiderverse_gwen/ \
--output_folder training_results/ood \
--run_name ood \
--num_pti_steps 240 \
--use_fa \
--center_sigma 0.0
Afterwards, editing is performed the same way:
python edit_video.py --input_folder datasets/ood_spiderverse_gwen/ \
--output_folder edits/ood \
--run_name ood \
--edit_name smile \
--edit_range 2 2 1
out.mp4
python edit_video.py --input_folder datasets/ood_spiderverse_gwen/ \
--output_folder edits/ood \
--run_name ood \
--edit_type styleclip_global \
--edit_range 10 10 1 \
--edit_name lipstick \
--target_class 'Face with lipstick'
out.mp4
StyleGAN2-ada model and implementation:
https://github.com/NVlabs/stylegan2-ada-pytorch
Copyright © 2021, NVIDIA Corporation.
Nvidia Source Code License https://nvlabs.github.io/stylegan2-ada-pytorch/license.html
PTI implementation:
https://github.com/danielroich/PTI
Copyright (c) 2021 Daniel Roich
License (MIT) https://github.com/danielroich/PTI/blob/main/LICENSE
LPIPS model and implementation:
https://github.com/richzhang/PerceptualSimilarity
Copyright (c) 2020, Sou Uchida
License (BSD 2-Clause) https://github.com/richzhang/PerceptualSimilarity/blob/master/LICENSE
e4e model and implementation:
https://github.com/omertov/encoder4editing
Copyright (c) 2021 omertov
License (MIT) https://github.com/omertov/encoder4editing/blob/main/LICENSE
StyleCLIP model and implementation:
https://github.com/orpatashnik/StyleCLIP
Copyright (c) 2021 orpatashnik
License (MIT) https://github.com/orpatashnik/StyleCLIP/blob/main/LICENSE
StyleGAN2 Distillation for Feed-forward Image Manipulation - for editing directions:
https://github.com/EvgenyKashin/stylegan2-distillation
Copyright (c) 2019, Yandex LLC
License (Creative Commons NonCommercial) https://github.com/EvgenyKashin/stylegan2-distillation/blob/master/LICENSE
face-alignment Library:
https://github.com/1adrianb/face-alignment
Copyright (c) 2017, Adrian Bulat
License (BSD 3-Clause License) https://github.com/1adrianb/face-alignment/blob/master/LICENSE
face-parsing.PyTorch:
https://github.com/zllrunning/face-parsing.PyTorch
Copyright (c) 2019 zll
License (MIT) https://github.com/zllrunning/face-parsing.PyTorch/blob/master/LICENSE
If you make use of our work, please cite our paper:
@misc{tzaban2022stitch,
title={Stitch it in Time: GAN-Based Facial Editing of Real Videos},
author={Rotem Tzaban and Ron Mokady and Rinon Gal and Amit H. Bermano and Daniel Cohen-Or},
year={2022},
eprint={2201.08361},
archivePrefix={arXiv},
primaryClass={cs.CV}
}