Teaser Video: Top row denotes source images, and the leftmost column denotes the reference images from which the style is being transferred.
- annotated paper
- paper implementation codes
- face alignment model from paper
- resnet model for attribute classification
- helper functions for latent interpolation
- style codes with predicted attributes for all images
- trained svm boudaries for selected attributed
- app using dash and flask
- celeb_a dataset (male, female) - full, sample
- celeb_a attribute dataset - link
- pretrained models - starganv2, classifier
- Session 1 recording - Intro to GANs, Understanding the paper.
- Session 2 recording part 1 - Dataloaders and basic model blocks, part 2 - complete model, training and validation helper functions.
- Session 3 recording - training large models on VMs, possible improvements, deploying app using dash and flask, interesting intuitions and manipulating latent space.
Interpolation with reference images
Interpolation of latent space to manipulate certain attributes
Masks generated using masking model to retain features like eyes, nose and mouth of the source image
Demon-like images during cross domain interpolation
Retention of domain characteristics irrespective of reference image
Well seperated domain latent spaces
Grad Cam to visualise activations to understand each feature of the latent space
