Date: October 2025.
Source: Proceedings of the IEEE/CVF International Conference on Computer Vision 2025 (pp. 10196-10206). ICCV, Honolulu, Hawai’i.
Research Summary: ImHead addresses a key limitation of 3D Morphable Models (3DMMs), the standard approach for modeling and generating expressive 3D avatars, whose strict topology and linear nature struggle to represent complex full-head shapes. Building on the advent of deep implicit functions, ImHead introduces a novel implicit 3DMM that models expressive 3D head avatars and enables localized editing of facial features. Rather than dividing the latent space into separate local components, as previous methods did, ImHead retains a single compact identity space paired with an intermediate region-specific latent representation, enabling local edits without the added latent size. To train ImHead, the team curated a large-scale real-world face dataset of over 4,000 distinct identities, a step toward large-scale 3D head modeling, and the model outperforms previous approaches at representing diverse identities and expressions while allowing interpretable, localized 3D face manipulation.
3dMD’s Role: ImHead uses real-world 3dMD face data captured originally for the MimicME project on a 3dMDface.t system running at 10fps at the Science Museum, London. ImHead, which includes 4,000 3D scans including over 600 children under 12 and more than 100 adults over 60, is a direct downstream application using MimicME’s real-world 3dMD data.
Article: ImHead: A Large-scale Implicit Morphable Model for Localized Head Modeling.
Authors: RA Potamias, S Galanakis, J Deng, A Papaioannou, S Zafeiriou. Faculty of Engineering, Department of Computing, Imperial College London.