This is the official repository accompanying the BMVC 2022 paper:
$S^2$-Flow: Joint Semantic and Style Editing of Facial Images
K. Singh, S. Schaub-Meyer, and S. Roth.

Figure: $S^2$-Flow is capable of applying semantic (top right), style (bottom right), and joint semantic and style edits (left) for facial images while preserving both identity and realism.

The high-quality images yielded by generative adversarial networks (GANs) have motivated investigations into their application for image editing. However, GANs are often limited in the control they provide for performing specific edits. One of the principal challenges is the entangled latent space of GANs, which is not directly suitable for performing independent and detailed edits. Recent editing methods allow for either controlled style edits or controlled semantic edits. In addition, methods that use semantic masks to edit images have difficulty preserving the identity and are unable to perform controlled style edits. We propose a method to disentangle a GAN’s latent space into semantic and style spaces, enabling controlled semantic and style edits for face images independently within the same framework. To achieve this, we design an encoder-decoder based network architecture $(S^2$-Flow), which incorporates two proposed inductive biases. We show the suitability of $S^2$-Flow quantitatively and qualitatively by performing various semantic and style edits.

[Paper] | [Video] | [Poster]

Clone this repo.

git clone https://github.com/visinf/s2-flow.git
cd s2-flow/

This code requires PyTorch, TensorFlow, Torchdiffeq, and Python 3+. Please install dependencies by

conda env create --file environment.yaml
python setup.py install

After this, go to this google drive, and

• Download experiments and place them under ./.
• Download and unzip data.zip and place it under ./.
• Download and unzip deeplab_model.zip and place it under ./.

python ./tools/run_semantic_edits.py --exp_path ./data/yaw/minus_20

The results are stored in a results/cs3_full folder in exp_path. You can create your semantic edits by uploading your edited grey and coloured semantic mask.

python ./tools/run_sequential_semantic_edit.py

The results for sequential changing the smile and hairstyle are stored in a folder named results_sequential/cs3_full folder.

python ./tools/run_style_editing.py

The results for hair colour change from red to black are stored in a folder named style_editing.

This experiment shows conditional generation results and style and style disentanglement results.

python ./tools/run_conditional.py

Conditional generation results named cond_gen.png are stored under ./results/cond_test folder. The disentanglement results are also stored in results folder, random_imgs_segmap_fixed_<?> shows results when the segmap is fixed and random_imgs_style_fixed_<?> shows results when the style is fixed.

If you find our work useful in your research please consider citing our papers:

@InProceedings{Singh:2022:SFJ,
    title   = {$S^2$-Flow: Joint Semantic and Style Editing of Facial Images}
    author  = {Krishnakant Singh and Simone Schaub-Meyer and Stefan Roth},
    booktitle = {British Machine Vision Conference (BMVC)},
    year = {2022}                         
}

This implementation builds upon the awesome work done by Abdal et al. StyleFlow. We would also like to thank the contributors of the following repositories for using parts of their publicly available code:

• StyleGAN2
• torchdiffeq
• PointFlow