3D Anime Faces Generation from Single Images by HoloGAN

The generation of 3D anime faces two critical challenges: limited 3D dataset availability and copyright constraints. This study proposes an innovative approach using a holographic generative adversarial network (HoloGAN), a generative adversarial network (GAN) that produces 3D-consistent images from 2D data. By training the model on an anime-styled adaptation of the CelebA datasets, we demonstrate high-quality, multi-view anime face generation without explicit 3D supervision. The method synthesizes consistent facial features across diverse viewpoints, addressing multi-view and rotational coherence challenges in 3D facial synthesis. Experiments validate the HoloGAN-based framework's capability to produce visually coherent and aesthetically appealing anime-style faces from various angles. This approach offers a viable alternative to 3D data-dependent methods and circumvents legal risks regarding dataset availability and copyright.