Skip to content
Book Open access

SelfieAvatar: High-Fidelity Animatable 3D Head Avatars from In-the-Wild Selfies

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 61 references

TL;DR

A Diffusion Transformer-based multi-view rectifier that canonicalizes the input selfies to predefined viewpoints while normalizing them to a shared facial expression is employed, which enables consistent multi-view facial animation with fine-grained expression preservation, including challenging motions such as tongue movement.

Abstract

We present SelfieAvatar, a practical system for generating high-fidelity, animatable 3D head avatars from three casually captured, in-the-wild selfies, without requiring camera calibration, fixed facial expressions, or 3D morphable model (3DMM) fitting. While multiple selfies provide stronger geometric cues than single-image input, they also introduce significant challenges due to unknown camera poses and inconsistent facial expressions across views. To address these issues, we employ a Diffusion Transformer (DiT)-based multi-view rectifier that canonicalizes the input selfies to predefined viewpoints while normalizing them to a shared facial expression. Based on these rectified canonical views, we use a feed-forward Gaussian splatting reconstruction model that produces detailed and geometrically consistent 3D head avatars. Furthermore, we propose a 3DMM-free 3D avatar reenactment framework based on latent diffusion, in which a compact motion embedding extracted from a driving image controls facial expression via cross-attention, and multi-view consistency modules enforce coherent animation across viewpoints. This design enables consistent multi-view facial animation with fine-grained expression preservation, including challenging motions such as tongue movement. Extensive experiments demonstrate that SelfieAvatar achieves superior reconstruction quality, view consistency, and animation realism compared to existing methods, highlighting its robustness and practicality for real-world 3D avatar creation from sparse casual inputs.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.