Mask-Guided Discovery of Semantic Manifolds in Generative Models
Mengyu Yang, David Rokeby, Xavier Snelgrove
BMO Lab for Creative Research, University of Toronto
in Workshop on Machine Learning for Creativity and Design (NeurIPS), 2020

A mask-guided, optimization based approach to learning semantic manifolds in StyleGAN2. Given an initial latent vector, this method finds additional vectors corresponding to the same image but with changes localized within the mask region.

Starting with an initial generated image and its corresponding latent vector (orange in the figure above), we define a rectangular mask region M over the image, such as around the mouth. The manifold we seek contains images and their latent vectors (red) that have primarily changed in the mask region but not in the rest of the image. We define this manifold as the minima of a function that measures the distance between the initial reference image and another generated image. Within this function, the non-mask region of the images are directly compared while the difference between the mask regions of both images is offset by an adjustable parameter, which leads to the function being minimal when the mask region has changed by that factor.

To create smoothly varying animations that explore this manifold, we implement a "spring" loss with an adjustable rest length term (green connectors in above figure) between vectors to encourage neighbours to be similar but not by too much. Higher-order “stiffener” springs (blue connectors) also connect to vectors that are further apart to encourage the entire length of vectors to have minimal curvature.

First, download the converted StyleGAN2 FFHQ checkpoint for its PyTorch implementation here and put it in the same directory as the rest of the code.

Dependencies:

After cloning the repo, run:

git submodule init; git submodule update

The dependencies can be installed as follows:

pip install -r requirements.txt

Note that CUDA 10.1/10.2 is also required to run the code.

python run.py

This method contains different arguments for hyperparameters controlling the weights of each loss term, the mask offset value, and the primary spring loss distance. Additional arguments also specify items such as the mask region, number of output vectors, and optimization epochs. Defaults have been set for all these arguments so the method can be easily used.

To see a list of all arguments with explanations:

python run.py -h

Under the default arguments, once the optimization finishes, the resulting latent vectors will be saved in file {exp_name}_latents.pt, the corresponding images will be saved in a new directory called {exp_name}_imgs, and a plot of the UMAP projection (https://arxiv.org/abs/1802.03426) of the latent vectors will be saved as {exp_name}_UMAP.png. {exp_name} is by default set to masked_gan_exp but this can be changed by:

python run.py --exp_name NEW_EXPERIMENT_NAME

(Below) Animation created from images generated using a mask region around the mouth. The mask offset is slowly increased as the animation progresses.

(Below) Combining the resulting latent directions from three mask regions using the same face. The bottom right animation is the combined result while the other 3 animations are the results of the individual experiments.

(Below) Taking the original result from an experiment as shown in the top left animation and applying the latent directions to other faces.

(Below) A UMAP projection of output latent vectors.

@inproceedings{yang2020mask-guided,
  title={Mask-Guided Discovery of Semantic Manifolds in Generative Models},
  author={Yang, Mengyu and Rokeby, David and Snelgrove, Xavier},
  year={2020},
  booktitle={4th Workshop on Machine Learning for Creativity and Design at NeurIPS 2020}
}