Skip to content

HVossi92/AI-driven-animations-of-2D-rasterized-graphics

Repository files navigation

AI-driven-animations-of-2D-rasterized-graphics

Master thesis: AI-driven animations of 2D rasterized graphics

Abstract

Manual keyframe animation is an integral part of the post-production process for 2D animations. Despite this, little research has been undertaken to utilize advances in artificial intelligence to improve this process, which still takes considerable user effort.

Inspired by recent progress in image manipulation and time series prediction through machine learning, this thesis examines the possibility of automatically creating animations between two rasterized graphics. The focus lies on small black and white 2D icon animations.

To this end, I create and compare different neural network architectures and present a multi-layered process built upon a convolutional neural network, called akaNet, for predicting the frames between a given input and output graphic. This approach gets augmented by a novel workflow using recursive triads for improved time series prediction results.

Contents