MegaPortraits: One-shot Megapixel Neural Head Avatars

• DOI: 10.48550/arxiv.2207.07621
• Publication Date: 2022-01-01
• Authors: Nikita Drobyshev; Jenya Chelishev; Taras Khakhulin; Aleksei Ivakhnenko; Victor Lempitsky; Egor Zakharov
• Keywords: Avatar; Leverage (statistics)

In this work, we advance the neural head avatar technology to megapixel resolution while focusing on particularly challenging task of cross-driving synthesis, i.e., when appearance driving image is substantially different from animated source image. We propose a set new architectures and training methods that can leverage both medium-resolution video data high-resolution achieve desired levels rendered quality generalization novel views motion. demonstrate suggested produce convincing avatars, outperforming competitors in scenario. Lastly, show how trained model be distilled into lightweight student which runs real-time locks identities avatars several dozens pre-defined images. Real-time operation identity lock are essential for many practical applications systems.