Experimentation on image-to-image translation with pix2pixHD. The goal/ hypothesis is to be able to convert face sketches to synthesized, stylistically realistic images.
For this model I am not training with image labels, just using sketch lines as the input training data set, with the intension to use this one for straight forward drawing input. For model training, I used I used the first 6,000 images with 512^2 resolution, with sketch + photo paired set as the training data set.
Dataset: Flickr-Faces-HQ Dataset (FFHQ)
ML Model: pix2pixHD
To make an interactive ML web app prototype, I used Streamlit (front end), Colab (as the backend), and created a temporary URL using ngrok. See the video documentation of the ML web app below.
Result with Face Dataset
Input with color masks
Input (Egon Schiele)
Output
Output
Result with Shoes Dataset
Sketch input
Sketch input
Sketch input
Output
Output
Output