Comparison study of image transformations via latent space vector manipulation against standard image transformation models like pix-to-pix/CycleGAN
Please install the following python libraries
numpy
tensorflow 2.0 +
keras
- Data Scraping - Face collection from comics
- Anime Face Generation - GAN & Style-GAN based generation
- Facial Expression Change - latent embedding hyperplane transformation vs trained image transformation model
This personal project was mainly a study to compare the two pipeline results for image transformation models using GAN. Typically, Image transformation models use input and output samples which may not be paired. As far as use of latent vectors are concerned, they are used in case studies to investigate how the model interprets the input distribution and used to perform transformations of an image(e.g from a gender / age/ race of a person to another) via learned hyperplane translations of the latent vector. Thus this project seeks to compare the qualitative results of the direct appraoch versus this indirect via optimization using a classifier model. This can allow greater user control of the transformations should the results be promising.
Anime_face_collector.py is a file that simply scans for an image and outputs crops on the anime faces. It has a good success rate though not exactly perfect anime-facegan.py contains the baseline training code for generative purposes a2.py contains all the training and prediction code variations, indicated by comments, of anime-facegan.py experiments.py is the most important file. It contains various optimization schema and models for this process. The output folder contains some sample results from running the optimization functionality