This is an interactive image editing tool based on “Sparse, Smart Contours to Represent and Edit Images," CVPR '18 by Dekel, et al. The idea is to represent an image by a sparse set of contours + gradient information, and train a deep generative based model to invert it, i.e., reconstruct the image with high fidelity. The model learns to synthesize texture, details and fine structures in regions where no input information is provided, while gaining semantic knowledge about the training data. Image pixel manipulation is then driven by editing the underlying contours.
In this demo, editing is performed by selecting a bunch of contours (“select” button) and moving them around ("move" button), copying them ("copy" button), or simply removing contours (“erase” button). After editing the contours click “reconstruct” to invert the new, modified contours back into pixels.
The contour2im model used for the demo was trained on the VGG face dataset and implemented in TensorFlow.