AlteredAvatar: Stylizing Dynamic 3D Avatars with Fast Style Adaptation

TL;DR

AlteredAvatar is a meta-learning based method that enables rapid and high-quality stylization of dynamic 3D avatars from various style inputs, balancing speed, flexibility, and visual consistency.

Contribution

It introduces a novel meta-learning framework that allows quick adaptation of 3D avatar styles from text or images, outperforming existing slow optimization and less flexible feed-forward methods.

Findings

Achieves fast style adaptation within a few update steps.

Maintains visual consistency across views and expressions.

Balances speed, quality, and flexibility effectively.

Abstract

This paper presents a method that can quickly adapt dynamic 3D avatars to arbitrary text descriptions of novel styles. Among existing approaches for avatar stylization, direct optimization methods can produce excellent results for arbitrary styles but they are unpleasantly slow. Furthermore, they require redoing the optimization process from scratch for every new input. Fast approximation methods using feed-forward networks trained on a large dataset of style images can generate results for new inputs quickly, but tend not to generalize well to novel styles and fall short in quality. We therefore investigate a new approach, AlteredAvatar, that combines those two approaches using the meta-learning framework. In the inner loop, the model learns to optimize to match a single target style well; while in the outer loop, the model learns to stylize efficiently across many styles. After training, AlteredAvatar learns an initialization that can quickly adapt within a small number of update steps to a novel style, which can be given using texts, a reference image, or a combination of both. We show that AlteredAvatar can achieve a good balance between speed, flexibility and quality, while maintaining consistency across a wide range of novel views and facial expressions.