Faithful single image face reconstruction using GAN inversion
File(s)
Accepted version
Author(s)
Luo, Yiming
Ghosh, Abhijeet
Type
Conference Paper
Abstract
Three-dimensional generative adversarial networks (3D GANs) enable high-fidelity and view-consistent image synthesis for applications such as digital humans and avatars. However, existing inversion methods face key limitations. Encoder-based approaches, though efficient, often lose fine details and degrade under challenging poses. Optimization-based methods, while flexible, lack geometry supervision and produce distortions in novel views. Symmetry priors partly address these issues but restrict generalization and rely on fixed pose estimators. Moreover, poseonditioned 3D GANs assume known pose distributions, which may misalign with real data and misguide geometry learning. To overcome these challenges, we propose an optimization-based 3D GAN inversion framework built upon a pose-free generator that implicitly learns the pose distribution. This removes the need for pose annotations or rigid priors, while reformulating inversion as a latent-only optimization problem. We further introduce geometry-aware regularization to stabilize latent optimization and improve view consistency. Experiments on challenging benchmarks demonstrate that our approach achieves more faithful reconstructions, geometrically consistent novel views, and robust identity preservation under extreme poses and asymmetric structures compared to prior methods.
Date Acceptance
2026-02-06
Copyright Statement
Subject to copyright. This paper is embargoed until publication. Once published the author’s accepted manuscript will be made available under a CC-BY License in accordance with Imperial’s Research Publications Open Access policy (www.imperial.ac.uk/oa-policy).
Source
Computational Visual Media Conference (CVM)
Publication Status
Accepted
Start Date
2026-04-10
Finish Date
2026-04-12
Coverage Spatial
Seoul, South Korea